2018-03-07 08:17:48 +08:00
|
|
|
//===- BugReporter.cpp - Generate PathDiagnostics for bugs ----------------===//
|
2008-04-03 12:42:52 +08:00
|
|
|
//
|
2019-01-19 16:50:56 +08:00
|
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
2008-04-03 12:42:52 +08:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file defines BugReporter, a utility class for generating
|
2009-06-26 08:05:51 +08:00
|
|
|
// PathDiagnostics.
|
2008-04-03 12:42:52 +08:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2011-02-10 09:03:03 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/BugReporter/BugReporter.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/AST/Decl.h"
|
|
|
|
#include "clang/AST/DeclBase.h"
|
2012-01-28 20:06:22 +08:00
|
|
|
#include "clang/AST/DeclObjC.h"
|
2008-04-03 12:42:52 +08:00
|
|
|
#include "clang/AST/Expr.h"
|
2013-06-17 20:56:08 +08:00
|
|
|
#include "clang/AST/ExprCXX.h"
|
2009-03-28 04:55:39 +08:00
|
|
|
#include "clang/AST/ParentMap.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/AST/Stmt.h"
|
2013-06-17 20:56:08 +08:00
|
|
|
#include "clang/AST/StmtCXX.h"
|
2014-01-07 19:51:46 +08:00
|
|
|
#include "clang/AST/StmtObjC.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/Analysis/AnalysisDeclContext.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "clang/Analysis/CFG.h"
|
2016-12-22 22:48:52 +08:00
|
|
|
#include "clang/Analysis/CFGStmtMap.h"
|
2019-09-12 04:54:27 +08:00
|
|
|
#include "clang/Analysis/PathDiagnostic.h"
|
2008-04-03 12:42:52 +08:00
|
|
|
#include "clang/Analysis/ProgramPoint.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/Basic/LLVM.h"
|
|
|
|
#include "clang/Basic/SourceLocation.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "clang/Basic/SourceManager.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/AnalyzerOptions.h"
|
|
|
|
#include "clang/StaticAnalyzer/Core/BugReporter/BugReporterVisitors.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/BugReporter/BugType.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/Checker.h"
|
|
|
|
#include "clang/StaticAnalyzer/Core/CheckerManager.h"
|
2020-05-27 18:27:32 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/CheckerRegistryData.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/ExplodedGraph.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/ExprEngine.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/MemRegion.h"
|
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
|
2020-06-30 00:51:24 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/SMTConv.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/SVals.h"
|
|
|
|
#include "clang/StaticAnalyzer/Core/PathSensitive/SymbolManager.h"
|
|
|
|
#include "llvm/ADT/ArrayRef.h"
|
2008-06-18 13:34:07 +08:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "llvm/ADT/DenseSet.h"
|
|
|
|
#include "llvm/ADT/FoldingSet.h"
|
|
|
|
#include "llvm/ADT/None.h"
|
|
|
|
#include "llvm/ADT/Optional.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "llvm/ADT/SmallString.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2013-03-16 05:41:53 +08:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2020-04-07 01:32:16 +08:00
|
|
|
#include "llvm/ADT/StringExtras.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include "llvm/ADT/StringRef.h"
|
|
|
|
#include "llvm/ADT/iterator_range.h"
|
|
|
|
#include "llvm/Support/Casting.h"
|
|
|
|
#include "llvm/Support/Compiler.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
|
|
|
#include "llvm/Support/MemoryBuffer.h"
|
2012-12-04 17:13:33 +08:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2018-03-07 08:17:48 +08:00
|
|
|
#include <algorithm>
|
|
|
|
#include <cassert>
|
|
|
|
#include <cstddef>
|
|
|
|
#include <iterator>
|
2014-03-09 19:36:40 +08:00
|
|
|
#include <memory>
|
2009-03-13 07:41:59 +08:00
|
|
|
#include <queue>
|
2018-03-07 08:17:48 +08:00
|
|
|
#include <string>
|
|
|
|
#include <tuple>
|
|
|
|
#include <utility>
|
|
|
|
#include <vector>
|
2008-04-03 12:42:52 +08:00
|
|
|
|
|
|
|
using namespace clang;
|
2010-12-23 15:20:52 +08:00
|
|
|
using namespace ento;
|
2019-09-10 04:34:40 +08:00
|
|
|
using namespace llvm;
|
2008-04-03 12:42:52 +08:00
|
|
|
|
2014-04-22 11:17:02 +08:00
|
|
|
#define DEBUG_TYPE "BugReporter"
|
|
|
|
|
2013-03-16 05:41:53 +08:00
|
|
|
STATISTIC(MaxBugClassSize,
|
|
|
|
"The maximum number of bug reports in the same equivalence class");
|
|
|
|
STATISTIC(MaxValidBugClassSize,
|
|
|
|
"The maximum number of bug reports in the same equivalence class "
|
|
|
|
"where at least one report is valid (not suppressed)");
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
BugReporterVisitor::~BugReporterVisitor() = default;
|
2010-03-21 02:01:57 +08:00
|
|
|
|
2011-12-20 10:48:34 +08:00
|
|
|
void BugReporterContext::anchor() {}
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// PathDiagnosticBuilder and its associated routines and helper objects.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
/// A (CallPiece, node assiciated with its CallEnter) pair.
|
|
|
|
using CallWithEntry =
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
std::pair<PathDiagnosticCallPiece *, const ExplodedNode *>;
|
2019-08-14 04:42:48 +08:00
|
|
|
using CallWithEntryStack = SmallVector<CallWithEntry, 6>;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
|
|
|
/// Map from each node to the diagnostic pieces visitors emit for them.
|
|
|
|
using VisitorsDiagnosticsTy =
|
|
|
|
llvm::DenseMap<const ExplodedNode *, std::vector<PathDiagnosticPieceRef>>;
|
|
|
|
|
|
|
|
/// A map from PathDiagnosticPiece to the LocationContext of the inlined
|
|
|
|
/// function call it represents.
|
|
|
|
using LocationContextMap =
|
|
|
|
llvm::DenseMap<const PathPieces *, const LocationContext *>;
|
|
|
|
|
|
|
|
/// A helper class that contains everything needed to construct a
|
|
|
|
/// PathDiagnostic object. It does no much more then providing convenient
|
|
|
|
/// getters and some well placed asserts for extra security.
|
2019-08-14 04:42:48 +08:00
|
|
|
class PathDiagnosticConstruct {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
/// The consumer we're constructing the bug report for.
|
|
|
|
const PathDiagnosticConsumer *Consumer;
|
|
|
|
/// Our current position in the bug path, which is owned by
|
|
|
|
/// PathDiagnosticBuilder.
|
|
|
|
const ExplodedNode *CurrentNode;
|
|
|
|
/// A mapping from parts of the bug path (for example, a function call, which
|
|
|
|
/// would span backwards from a CallExit to a CallEnter with the nodes in
|
|
|
|
/// between them) with the location contexts it is associated with.
|
|
|
|
LocationContextMap LCM;
|
|
|
|
const SourceManager &SM;
|
|
|
|
|
|
|
|
public:
|
|
|
|
/// We keep stack of calls to functions as we're ascending the bug path.
|
|
|
|
/// TODO: PathDiagnostic has a stack doing the same thing, shouldn't we use
|
|
|
|
/// that instead?
|
2019-08-14 04:42:48 +08:00
|
|
|
CallWithEntryStack CallStack;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
/// The bug report we're constructing. For ease of use, this field is kept
|
|
|
|
/// public, though some "shortcut" getters are provided for commonly used
|
|
|
|
/// methods of PathDiagnostic.
|
|
|
|
std::unique_ptr<PathDiagnostic> PD;
|
|
|
|
|
|
|
|
public:
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticConstruct(const PathDiagnosticConsumer *PDC,
|
2019-09-10 04:34:40 +08:00
|
|
|
const ExplodedNode *ErrorNode,
|
|
|
|
const PathSensitiveBugReport *R);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
|
|
|
/// \returns the location context associated with the current position in the
|
|
|
|
/// bug path.
|
|
|
|
const LocationContext *getCurrLocationContext() const {
|
|
|
|
assert(CurrentNode && "Already reached the root!");
|
|
|
|
return CurrentNode->getLocationContext();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Same as getCurrLocationContext (they should always return the same
|
|
|
|
/// location context), but works after reaching the root of the bug path as
|
|
|
|
/// well.
|
|
|
|
const LocationContext *getLocationContextForActivePath() const {
|
|
|
|
return LCM.find(&PD->getActivePath())->getSecond();
|
|
|
|
}
|
|
|
|
|
|
|
|
const ExplodedNode *getCurrentNode() const { return CurrentNode; }
|
|
|
|
|
|
|
|
/// Steps the current node to its predecessor.
|
|
|
|
/// \returns whether we reached the root of the bug path.
|
|
|
|
bool ascendToPrevNode() {
|
|
|
|
CurrentNode = CurrentNode->getFirstPred();
|
|
|
|
return static_cast<bool>(CurrentNode);
|
|
|
|
}
|
|
|
|
|
|
|
|
const ParentMap &getParentMap() const {
|
|
|
|
return getCurrLocationContext()->getParentMap();
|
|
|
|
}
|
|
|
|
|
|
|
|
const SourceManager &getSourceManager() const { return SM; }
|
|
|
|
|
|
|
|
const Stmt *getParent(const Stmt *S) const {
|
|
|
|
return getParentMap().getParent(S);
|
|
|
|
}
|
|
|
|
|
|
|
|
void updateLocCtxMap(const PathPieces *Path, const LocationContext *LC) {
|
|
|
|
assert(Path && LC);
|
|
|
|
LCM[Path] = LC;
|
|
|
|
}
|
|
|
|
|
|
|
|
const LocationContext *getLocationContextFor(const PathPieces *Path) const {
|
|
|
|
assert(LCM.count(Path) &&
|
|
|
|
"Failed to find the context associated with these pieces!");
|
|
|
|
return LCM.find(Path)->getSecond();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool isInLocCtxMap(const PathPieces *Path) const { return LCM.count(Path); }
|
|
|
|
|
|
|
|
PathPieces &getActivePath() { return PD->getActivePath(); }
|
|
|
|
PathPieces &getMutablePieces() { return PD->getMutablePieces(); }
|
|
|
|
|
|
|
|
bool shouldAddPathEdges() const { return Consumer->shouldAddPathEdges(); }
|
2020-12-01 16:53:53 +08:00
|
|
|
bool shouldAddControlNotes() const {
|
|
|
|
return Consumer->shouldAddControlNotes();
|
|
|
|
}
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
bool shouldGenerateDiagnostics() const {
|
|
|
|
return Consumer->shouldGenerateDiagnostics();
|
|
|
|
}
|
|
|
|
bool supportsLogicalOpControlFlow() const {
|
|
|
|
return Consumer->supportsLogicalOpControlFlow();
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Contains every contextual information needed for constructing a
|
2019-08-14 06:03:08 +08:00
|
|
|
/// PathDiagnostic object for a given bug report. This class and its fields are
|
|
|
|
/// immutable, and passes a BugReportConstruct object around during the
|
|
|
|
/// construction.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
class PathDiagnosticBuilder : public BugReporterContext {
|
|
|
|
/// A linear path from the error node to the root.
|
|
|
|
std::unique_ptr<const ExplodedGraph> BugPath;
|
2019-08-14 06:03:08 +08:00
|
|
|
/// The bug report we're describing. Visitors create their diagnostics with
|
|
|
|
/// them being the last entities being able to modify it (for example,
|
|
|
|
/// changing interestingness here would cause inconsistencies as to how this
|
|
|
|
/// file and visitors construct diagnostics), hence its const.
|
2019-09-10 04:34:40 +08:00
|
|
|
const PathSensitiveBugReport *R;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
/// The leaf of the bug path. This isn't the same as the bug reports error
|
|
|
|
/// node, which refers to the *original* graph, not the bug path.
|
|
|
|
const ExplodedNode *const ErrorNode;
|
|
|
|
/// The diagnostic pieces visitors emitted, which is expected to be collected
|
|
|
|
/// by the time this builder is constructed.
|
|
|
|
std::unique_ptr<const VisitorsDiagnosticsTy> VisitorsDiagnostics;
|
|
|
|
|
|
|
|
public:
|
|
|
|
/// Find a non-invalidated report for a given equivalence class, and returns
|
|
|
|
/// a PathDiagnosticBuilder able to construct bug reports for different
|
|
|
|
/// consumers. Returns None if no valid report is found.
|
|
|
|
static Optional<PathDiagnosticBuilder>
|
2019-09-10 04:34:40 +08:00
|
|
|
findValidReport(ArrayRef<PathSensitiveBugReport *> &bugReports,
|
2019-08-20 10:15:50 +08:00
|
|
|
PathSensitiveBugReporter &Reporter);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
|
|
|
PathDiagnosticBuilder(
|
|
|
|
BugReporterContext BRC, std::unique_ptr<ExplodedGraph> BugPath,
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport *r, const ExplodedNode *ErrorNode,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
std::unique_ptr<VisitorsDiagnosticsTy> VisitorsDiagnostics);
|
|
|
|
|
|
|
|
/// This function is responsible for generating diagnostic pieces that are
|
|
|
|
/// *not* provided by bug report visitors.
|
|
|
|
/// These diagnostics may differ depending on the consumer's settings,
|
|
|
|
/// and are therefore constructed separately for each consumer.
|
|
|
|
///
|
|
|
|
/// There are two path diagnostics generation modes: with adding edges (used
|
|
|
|
/// for plists) and without (used for HTML and text). When edges are added,
|
|
|
|
/// the path is modified to insert artificially generated edges.
|
|
|
|
/// Otherwise, more detailed diagnostics is emitted for block edges,
|
|
|
|
/// explaining the transitions in words.
|
|
|
|
std::unique_ptr<PathDiagnostic>
|
|
|
|
generate(const PathDiagnosticConsumer *PDC) const;
|
|
|
|
|
|
|
|
private:
|
2019-09-12 04:54:17 +08:00
|
|
|
void updateStackPiecesWithMessage(PathDiagnosticPieceRef P,
|
|
|
|
const CallWithEntryStack &CallStack) const;
|
2019-08-14 04:42:48 +08:00
|
|
|
void generatePathDiagnosticsForNode(PathDiagnosticConstruct &C,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation &PrevLoc) const;
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
void generateMinimalDiagForBlockEdge(PathDiagnosticConstruct &C,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
BlockEdge BE) const;
|
|
|
|
|
|
|
|
PathDiagnosticPieceRef
|
2019-08-14 04:42:48 +08:00
|
|
|
generateDiagForGotoOP(const PathDiagnosticConstruct &C, const Stmt *S,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation &Start) const;
|
|
|
|
|
|
|
|
PathDiagnosticPieceRef
|
2019-08-14 04:42:48 +08:00
|
|
|
generateDiagForSwitchOP(const PathDiagnosticConstruct &C, const CFGBlock *Dst,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation &Start) const;
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticPieceRef
|
|
|
|
generateDiagForBinaryOP(const PathDiagnosticConstruct &C, const Stmt *T,
|
|
|
|
const CFGBlock *Src, const CFGBlock *DstC) const;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticLocation
|
|
|
|
ExecutionContinues(const PathDiagnosticConstruct &C) const;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticLocation
|
|
|
|
ExecutionContinues(llvm::raw_string_ostream &os,
|
|
|
|
const PathDiagnosticConstruct &C) const;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
const PathSensitiveBugReport *getBugReport() const { return R; }
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
} // namespace
|
|
|
|
|
2019-09-12 04:54:17 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Base implementation of stack hint generators.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
StackHintGenerator::~StackHintGenerator() = default;
|
|
|
|
|
|
|
|
std::string StackHintGeneratorForSymbol::getMessage(const ExplodedNode *N){
|
|
|
|
if (!N)
|
|
|
|
return getMessageForSymbolNotFound();
|
|
|
|
|
|
|
|
ProgramPoint P = N->getLocation();
|
|
|
|
CallExitEnd CExit = P.castAs<CallExitEnd>();
|
|
|
|
|
|
|
|
// FIXME: Use CallEvent to abstract this over all calls.
|
|
|
|
const Stmt *CallSite = CExit.getCalleeContext()->getCallSite();
|
|
|
|
const auto *CE = dyn_cast_or_null<CallExpr>(CallSite);
|
|
|
|
if (!CE)
|
|
|
|
return {};
|
|
|
|
|
|
|
|
// Check if one of the parameters are set to the interesting symbol.
|
|
|
|
unsigned ArgIndex = 0;
|
|
|
|
for (CallExpr::const_arg_iterator I = CE->arg_begin(),
|
|
|
|
E = CE->arg_end(); I != E; ++I, ++ArgIndex){
|
|
|
|
SVal SV = N->getSVal(*I);
|
|
|
|
|
|
|
|
// Check if the variable corresponding to the symbol is passed by value.
|
|
|
|
SymbolRef AS = SV.getAsLocSymbol();
|
|
|
|
if (AS == Sym) {
|
|
|
|
return getMessageForArg(*I, ArgIndex);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check if the parameter is a pointer to the symbol.
|
|
|
|
if (Optional<loc::MemRegionVal> Reg = SV.getAs<loc::MemRegionVal>()) {
|
|
|
|
// Do not attempt to dereference void*.
|
|
|
|
if ((*I)->getType()->isVoidPointerType())
|
|
|
|
continue;
|
|
|
|
SVal PSV = N->getState()->getSVal(Reg->getRegion());
|
|
|
|
SymbolRef AS = PSV.getAsLocSymbol();
|
|
|
|
if (AS == Sym) {
|
|
|
|
return getMessageForArg(*I, ArgIndex);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check if we are returning the interesting symbol.
|
|
|
|
SVal SV = N->getSVal(CE);
|
|
|
|
SymbolRef RetSym = SV.getAsLocSymbol();
|
|
|
|
if (RetSym == Sym) {
|
|
|
|
return getMessageForReturn(CE);
|
|
|
|
}
|
|
|
|
|
|
|
|
return getMessageForSymbolNotFound();
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string StackHintGeneratorForSymbol::getMessageForArg(const Expr *ArgE,
|
|
|
|
unsigned ArgIndex) {
|
|
|
|
// Printed parameters start at 1, not 0.
|
|
|
|
++ArgIndex;
|
|
|
|
|
|
|
|
return (llvm::Twine(Msg) + " via " + std::to_string(ArgIndex) +
|
|
|
|
llvm::getOrdinalSuffix(ArgIndex) + " parameter").str();
|
|
|
|
}
|
|
|
|
|
2012-02-29 07:06:21 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Diagnostic cleanup.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-10-26 06:07:10 +08:00
|
|
|
static PathDiagnosticEventPiece *
|
|
|
|
eventsDescribeSameCondition(PathDiagnosticEventPiece *X,
|
|
|
|
PathDiagnosticEventPiece *Y) {
|
|
|
|
// Prefer diagnostics that come from ConditionBRVisitor over
|
2016-10-05 16:19:49 +08:00
|
|
|
// those that came from TrackConstraintBRVisitor,
|
|
|
|
// unless the one from ConditionBRVisitor is
|
|
|
|
// its generic fallback diagnostic.
|
2012-10-26 06:07:10 +08:00
|
|
|
const void *tagPreferred = ConditionBRVisitor::getTag();
|
|
|
|
const void *tagLesser = TrackConstraintBRVisitor::getTag();
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-10-26 06:07:10 +08:00
|
|
|
if (X->getLocation() != Y->getLocation())
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
|
|
|
|
2012-10-26 06:07:10 +08:00
|
|
|
if (X->getTag() == tagPreferred && Y->getTag() == tagLesser)
|
2016-10-05 16:19:49 +08:00
|
|
|
return ConditionBRVisitor::isPieceMessageGeneric(X) ? Y : X;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-10-26 06:07:10 +08:00
|
|
|
if (Y->getTag() == tagPreferred && X->getTag() == tagLesser)
|
2016-10-05 16:19:49 +08:00
|
|
|
return ConditionBRVisitor::isPieceMessageGeneric(Y) ? X : Y;
|
2014-05-27 10:45:47 +08:00
|
|
|
|
|
|
|
return nullptr;
|
2012-10-26 06:07:10 +08:00
|
|
|
}
|
|
|
|
|
2012-10-27 00:02:36 +08:00
|
|
|
/// An optimization pass over PathPieces that removes redundant diagnostics
|
|
|
|
/// generated by both ConditionBRVisitor and TrackConstraintBRVisitor. Both
|
|
|
|
/// BugReporterVisitors use different methods to generate diagnostics, with
|
|
|
|
/// one capable of emitting diagnostics in some cases but not in others. This
|
|
|
|
/// can lead to redundant diagnostic pieces at the same point in a path.
|
|
|
|
static void removeRedundantMsgs(PathPieces &path) {
|
2012-10-26 06:07:10 +08:00
|
|
|
unsigned N = path.size();
|
|
|
|
if (N < 2)
|
|
|
|
return;
|
2012-10-27 00:02:36 +08:00
|
|
|
// NOTE: this loop intentionally is not using an iterator. Instead, we
|
|
|
|
// are streaming the path and modifying it in place. This is done by
|
|
|
|
// grabbing the front, processing it, and if we decide to keep it append
|
|
|
|
// it to the end of the path. The entire path is processed in this way.
|
2012-10-26 06:07:10 +08:00
|
|
|
for (unsigned i = 0; i < N; ++i) {
|
2017-01-06 01:26:53 +08:00
|
|
|
auto piece = std::move(path.front());
|
2012-10-26 06:07:10 +08:00
|
|
|
path.pop_front();
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-10-26 06:07:10 +08:00
|
|
|
switch (piece->getKind()) {
|
2016-10-07 23:23:02 +08:00
|
|
|
case PathDiagnosticPiece::Call:
|
2017-01-06 01:26:53 +08:00
|
|
|
removeRedundantMsgs(cast<PathDiagnosticCallPiece>(*piece).path);
|
2012-10-26 06:07:10 +08:00
|
|
|
break;
|
2016-10-07 23:23:02 +08:00
|
|
|
case PathDiagnosticPiece::Macro:
|
2017-01-06 01:26:53 +08:00
|
|
|
removeRedundantMsgs(cast<PathDiagnosticMacroPiece>(*piece).subPieces);
|
2012-10-26 06:07:10 +08:00
|
|
|
break;
|
2016-10-07 23:23:02 +08:00
|
|
|
case PathDiagnosticPiece::Event: {
|
2012-10-26 06:07:10 +08:00
|
|
|
if (i == N-1)
|
|
|
|
break;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (auto *nextEvent =
|
2014-07-05 11:08:06 +08:00
|
|
|
dyn_cast<PathDiagnosticEventPiece>(path.front().get())) {
|
2018-03-07 08:17:48 +08:00
|
|
|
auto *event = cast<PathDiagnosticEventPiece>(piece.get());
|
2012-10-26 06:07:10 +08:00
|
|
|
// Check to see if we should keep one of the two pieces. If we
|
|
|
|
// come up with a preference, record which piece to keep, and consume
|
|
|
|
// another piece from the path.
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *pieceToKeep =
|
|
|
|
eventsDescribeSameCondition(event, nextEvent)) {
|
|
|
|
piece = std::move(pieceToKeep == event ? piece : path.front());
|
2012-10-26 06:07:10 +08:00
|
|
|
path.pop_front();
|
|
|
|
++i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
case PathDiagnosticPiece::ControlFlow:
|
2016-10-07 23:23:02 +08:00
|
|
|
case PathDiagnosticPiece::Note:
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
case PathDiagnosticPiece::PopUp:
|
2016-10-07 23:23:02 +08:00
|
|
|
break;
|
2012-10-26 06:07:10 +08:00
|
|
|
}
|
2017-01-06 01:26:53 +08:00
|
|
|
path.push_back(std::move(piece));
|
2012-10-26 06:07:10 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-02-29 07:06:21 +08:00
|
|
|
/// Recursively scan through a path and prune out calls and macros pieces
|
|
|
|
/// that aren't needed. Return true if afterwards the path contains
|
2012-12-08 03:56:29 +08:00
|
|
|
/// "interesting stuff" which means it shouldn't be pruned from the parent path.
|
2019-08-14 04:42:48 +08:00
|
|
|
static bool removeUnneededCalls(const PathDiagnosticConstruct &C,
|
2019-09-10 04:34:40 +08:00
|
|
|
PathPieces &pieces,
|
|
|
|
const PathSensitiveBugReport *R,
|
2018-04-04 02:52:30 +08:00
|
|
|
bool IsInteresting = false) {
|
|
|
|
bool containsSomethingInteresting = IsInteresting;
|
2012-02-29 07:06:21 +08:00
|
|
|
const unsigned N = pieces.size();
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-02-29 07:06:21 +08:00
|
|
|
for (unsigned i = 0 ; i < N ; ++i) {
|
|
|
|
// Remove the front piece from the path. If it is still something we
|
|
|
|
// want to keep once we are done, we will push it back on the end.
|
2017-01-06 01:26:53 +08:00
|
|
|
auto piece = std::move(pieces.front());
|
2012-02-29 07:06:21 +08:00
|
|
|
pieces.pop_front();
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-03-01 08:05:06 +08:00
|
|
|
switch (piece->getKind()) {
|
|
|
|
case PathDiagnosticPiece::Call: {
|
2017-01-06 01:26:53 +08:00
|
|
|
auto &call = cast<PathDiagnosticCallPiece>(*piece);
|
2012-08-30 05:22:37 +08:00
|
|
|
// Check if the location context is interesting.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!removeUnneededCalls(
|
|
|
|
C, call.path, R,
|
|
|
|
R->isInteresting(C.getLocationContextFor(&call.path))))
|
2012-11-15 10:07:23 +08:00
|
|
|
continue;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-03-01 08:05:06 +08:00
|
|
|
containsSomethingInteresting = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case PathDiagnosticPiece::Macro: {
|
2017-01-06 01:26:53 +08:00
|
|
|
auto ¯o = cast<PathDiagnosticMacroPiece>(*piece);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!removeUnneededCalls(C, macro.subPieces, R, IsInteresting))
|
2012-03-01 08:05:06 +08:00
|
|
|
continue;
|
2012-02-29 07:06:21 +08:00
|
|
|
containsSomethingInteresting = true;
|
2012-03-01 08:05:06 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case PathDiagnosticPiece::Event: {
|
2017-01-06 01:26:53 +08:00
|
|
|
auto &event = cast<PathDiagnosticEventPiece>(*piece);
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-03-01 08:05:06 +08:00
|
|
|
// We never throw away an event, but we do throw it away wholesale
|
|
|
|
// as part of a path if we throw the entire path away.
|
2017-01-06 01:26:53 +08:00
|
|
|
containsSomethingInteresting |= !event.isPrunable();
|
2012-03-01 08:05:06 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case PathDiagnosticPiece::ControlFlow:
|
2016-10-07 23:23:02 +08:00
|
|
|
case PathDiagnosticPiece::Note:
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
case PathDiagnosticPiece::PopUp:
|
2016-10-07 23:23:02 +08:00
|
|
|
break;
|
2012-02-29 07:06:21 +08:00
|
|
|
}
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
pieces.push_back(std::move(piece));
|
2012-02-29 07:06:21 +08:00
|
|
|
}
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2012-02-29 07:06:21 +08:00
|
|
|
return containsSomethingInteresting;
|
|
|
|
}
|
|
|
|
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
/// Same logic as above to remove extra pieces.
|
|
|
|
static void removePopUpNotes(PathPieces &Path) {
|
|
|
|
for (unsigned int i = 0; i < Path.size(); ++i) {
|
|
|
|
auto Piece = std::move(Path.front());
|
|
|
|
Path.pop_front();
|
|
|
|
if (!isa<PathDiagnosticPopUpPiece>(*Piece))
|
|
|
|
Path.push_back(std::move(Piece));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-05-25 05:43:11 +08:00
|
|
|
/// Returns true if the given decl has been implicitly given a body, either by
|
|
|
|
/// the analyzer or by the compiler proper.
|
|
|
|
static bool hasImplicitBody(const Decl *D) {
|
|
|
|
assert(D);
|
|
|
|
return D->isImplicit() || !D->hasBody();
|
|
|
|
}
|
|
|
|
|
2012-12-08 03:56:29 +08:00
|
|
|
/// Recursively scan through a path and make sure that all call pieces have
|
2015-09-08 11:50:52 +08:00
|
|
|
/// valid locations.
|
2014-05-27 10:45:47 +08:00
|
|
|
static void
|
|
|
|
adjustCallLocations(PathPieces &Pieces,
|
|
|
|
PathDiagnosticLocation *LastCallLocation = nullptr) {
|
2018-03-07 08:17:48 +08:00
|
|
|
for (const auto &I : Pieces) {
|
|
|
|
auto *Call = dyn_cast<PathDiagnosticCallPiece>(I.get());
|
2012-12-08 03:56:29 +08:00
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (!Call)
|
2012-12-08 03:56:29 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
if (LastCallLocation) {
|
2013-05-25 05:43:11 +08:00
|
|
|
bool CallerIsImplicit = hasImplicitBody(Call->getCaller());
|
|
|
|
if (CallerIsImplicit || !Call->callEnter.asLocation().isValid())
|
2012-12-08 03:56:29 +08:00
|
|
|
Call->callEnter = *LastCallLocation;
|
2013-05-25 05:43:11 +08:00
|
|
|
if (CallerIsImplicit || !Call->callReturn.asLocation().isValid())
|
2012-12-08 03:56:29 +08:00
|
|
|
Call->callReturn = *LastCallLocation;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Recursively clean out the subclass. Keep this call around if
|
|
|
|
// it contains any informative diagnostics.
|
|
|
|
PathDiagnosticLocation *ThisCallLocation;
|
2013-01-22 02:28:30 +08:00
|
|
|
if (Call->callEnterWithin.asLocation().isValid() &&
|
2013-05-25 05:43:11 +08:00
|
|
|
!hasImplicitBody(Call->getCallee()))
|
2012-12-08 03:56:29 +08:00
|
|
|
ThisCallLocation = &Call->callEnterWithin;
|
|
|
|
else
|
|
|
|
ThisCallLocation = &Call->callEnter;
|
|
|
|
|
|
|
|
assert(ThisCallLocation && "Outermost call has an invalid location");
|
|
|
|
adjustCallLocations(Call->path, ThisCallLocation);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-10-17 01:45:35 +08:00
|
|
|
/// Remove edges in and out of C++ default initializer expressions. These are
|
|
|
|
/// for fields that have in-class initializers, as opposed to being initialized
|
|
|
|
/// explicitly in a constructor or braced list.
|
|
|
|
static void removeEdgesToDefaultInitializers(PathPieces &Pieces) {
|
|
|
|
for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *C = dyn_cast<PathDiagnosticCallPiece>(I->get()))
|
2013-10-17 01:45:35 +08:00
|
|
|
removeEdgesToDefaultInitializers(C->path);
|
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *M = dyn_cast<PathDiagnosticMacroPiece>(I->get()))
|
2013-10-17 01:45:35 +08:00
|
|
|
removeEdgesToDefaultInitializers(M->subPieces);
|
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *CF = dyn_cast<PathDiagnosticControlFlowPiece>(I->get())) {
|
2013-10-17 01:45:35 +08:00
|
|
|
const Stmt *Start = CF->getStartLocation().asStmt();
|
|
|
|
const Stmt *End = CF->getEndLocation().asStmt();
|
2021-10-20 23:43:31 +08:00
|
|
|
if (isa_and_nonnull<CXXDefaultInitExpr>(Start)) {
|
2013-10-17 01:45:35 +08:00
|
|
|
I = Pieces.erase(I);
|
|
|
|
continue;
|
2021-10-20 23:43:31 +08:00
|
|
|
} else if (isa_and_nonnull<CXXDefaultInitExpr>(End)) {
|
2014-03-02 20:20:24 +08:00
|
|
|
PathPieces::iterator Next = std::next(I);
|
2013-10-17 01:45:35 +08:00
|
|
|
if (Next != E) {
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *NextCF =
|
|
|
|
dyn_cast<PathDiagnosticControlFlowPiece>(Next->get())) {
|
2013-10-17 01:45:35 +08:00
|
|
|
NextCF->setStartLocation(CF->getStartLocation());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
I = Pieces.erase(I);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
I++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-06-07 06:02:58 +08:00
|
|
|
/// Remove all pieces with invalid locations as these cannot be serialized.
|
|
|
|
/// We might have pieces with invalid locations as a result of inlining Body
|
|
|
|
/// Farm generated functions.
|
|
|
|
static void removePiecesWithInvalidLocations(PathPieces &Pieces) {
|
2013-06-07 06:32:11 +08:00
|
|
|
for (PathPieces::iterator I = Pieces.begin(), E = Pieces.end(); I != E;) {
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *C = dyn_cast<PathDiagnosticCallPiece>(I->get()))
|
2013-06-07 06:02:58 +08:00
|
|
|
removePiecesWithInvalidLocations(C->path);
|
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *M = dyn_cast<PathDiagnosticMacroPiece>(I->get()))
|
2013-06-07 06:02:58 +08:00
|
|
|
removePiecesWithInvalidLocations(M->subPieces);
|
|
|
|
|
|
|
|
if (!(*I)->getLocation().isValid() ||
|
|
|
|
!(*I)->getLocation().asLocation().isValid()) {
|
2013-06-07 06:32:11 +08:00
|
|
|
I = Pieces.erase(I);
|
2013-06-07 06:02:58 +08:00
|
|
|
continue;
|
|
|
|
}
|
2013-06-07 06:32:11 +08:00
|
|
|
I++;
|
2013-06-07 06:02:58 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticLocation PathDiagnosticBuilder::ExecutionContinues(
|
|
|
|
const PathDiagnosticConstruct &C) const {
|
2019-09-12 04:54:21 +08:00
|
|
|
if (const Stmt *S = C.getCurrentNode()->getNextStmtForDiagnostics())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
return PathDiagnosticLocation(S, getSourceManager(),
|
|
|
|
C.getCurrLocationContext());
|
2009-03-28 04:55:39 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
return PathDiagnosticLocation::createDeclEnd(C.getCurrLocationContext(),
|
2011-09-17 03:18:30 +08:00
|
|
|
getSourceManager());
|
2009-03-13 02:41:53 +08:00
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticLocation PathDiagnosticBuilder::ExecutionContinues(
|
|
|
|
llvm::raw_string_ostream &os, const PathDiagnosticConstruct &C) const {
|
2008-05-07 02:11:09 +08:00
|
|
|
// Slow, but probably doesn't matter.
|
2009-02-24 06:44:26 +08:00
|
|
|
if (os.str().empty())
|
|
|
|
os << ' ';
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const PathDiagnosticLocation &Loc = ExecutionContinues(C);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-03-28 04:55:39 +08:00
|
|
|
if (Loc.asStmt())
|
2009-02-24 06:44:26 +08:00
|
|
|
os << "Execution continues on line "
|
2011-07-26 05:09:52 +08:00
|
|
|
<< getSourceManager().getExpansionLineNumber(Loc.asLocation())
|
2009-05-07 05:39:49 +08:00
|
|
|
<< '.';
|
2009-12-05 04:34:31 +08:00
|
|
|
else {
|
|
|
|
os << "Execution jumps to the end of the ";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const Decl *D = C.getCurrLocationContext()->getDecl();
|
2009-12-05 04:34:31 +08:00
|
|
|
if (isa<ObjCMethodDecl>(D))
|
|
|
|
os << "method";
|
|
|
|
else if (isa<FunctionDecl>(D))
|
|
|
|
os << "function";
|
|
|
|
else {
|
|
|
|
assert(isa<BlockDecl>(D));
|
|
|
|
os << "anonymous block";
|
|
|
|
}
|
|
|
|
os << '.';
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-03-13 02:41:53 +08:00
|
|
|
return Loc;
|
2008-05-07 02:11:09 +08:00
|
|
|
}
|
|
|
|
|
2013-06-04 06:59:48 +08:00
|
|
|
static const Stmt *getEnclosingParent(const Stmt *S, const ParentMap &PM) {
|
2009-05-15 09:50:15 +08:00
|
|
|
if (isa<Expr>(S) && PM.isConsumedExpr(cast<Expr>(S)))
|
2013-06-04 06:59:48 +08:00
|
|
|
return PM.getParentIgnoreParens(S);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-05-15 09:50:15 +08:00
|
|
|
const Stmt *Parent = PM.getParentIgnoreParens(S);
|
2013-06-04 06:59:48 +08:00
|
|
|
if (!Parent)
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-06-04 06:59:48 +08:00
|
|
|
switch (Parent->getStmtClass()) {
|
|
|
|
case Stmt::ForStmtClass:
|
|
|
|
case Stmt::DoStmtClass:
|
|
|
|
case Stmt::WhileStmtClass:
|
|
|
|
case Stmt::ObjCForCollectionStmtClass:
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
case Stmt::CXXForRangeStmtClass:
|
2013-06-04 06:59:48 +08:00
|
|
|
return Parent;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
2009-05-15 09:50:15 +08:00
|
|
|
}
|
|
|
|
|
2013-06-04 06:59:48 +08:00
|
|
|
static PathDiagnosticLocation
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
getEnclosingStmtLocation(const Stmt *S, const LocationContext *LC,
|
|
|
|
bool allowNestedContexts = false) {
|
2013-06-04 06:59:48 +08:00
|
|
|
if (!S)
|
2018-03-07 08:17:48 +08:00
|
|
|
return {};
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const SourceManager &SMgr = LC->getDecl()->getASTContext().getSourceManager();
|
|
|
|
|
|
|
|
while (const Stmt *Parent = getEnclosingParent(S, LC->getParentMap())) {
|
2009-03-28 11:37:59 +08:00
|
|
|
switch (Parent->getStmtClass()) {
|
2009-04-01 14:13:56 +08:00
|
|
|
case Stmt::BinaryOperatorClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *B = cast<BinaryOperator>(Parent);
|
2009-04-01 14:13:56 +08:00
|
|
|
if (B->isLogicalOp())
|
2013-06-04 06:59:53 +08:00
|
|
|
return PathDiagnosticLocation(allowNestedContexts ? B : S, SMgr, LC);
|
2009-04-01 14:13:56 +08:00
|
|
|
break;
|
2009-09-09 23:08:12 +08:00
|
|
|
}
|
2009-03-28 11:37:59 +08:00
|
|
|
case Stmt::CompoundStmtClass:
|
|
|
|
case Stmt::StmtExprClass:
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-03-28 12:08:14 +08:00
|
|
|
case Stmt::ChooseExprClass:
|
|
|
|
// Similar to '?' if we are referring to condition, just have the edge
|
|
|
|
// point to the entire choose expression.
|
2013-06-04 06:59:53 +08:00
|
|
|
if (allowNestedContexts || cast<ChooseExpr>(Parent)->getCond() == S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(Parent, SMgr, LC);
|
2009-03-28 12:08:14 +08:00
|
|
|
else
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2011-02-17 18:25:35 +08:00
|
|
|
case Stmt::BinaryConditionalOperatorClass:
|
2009-03-28 12:08:14 +08:00
|
|
|
case Stmt::ConditionalOperatorClass:
|
|
|
|
// For '?', if we are referring to condition, just have the edge point
|
|
|
|
// to the entire '?' expression.
|
2013-06-04 06:59:53 +08:00
|
|
|
if (allowNestedContexts ||
|
|
|
|
cast<AbstractConditionalOperator>(Parent)->getCond() == S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(Parent, SMgr, LC);
|
2009-03-28 12:08:14 +08:00
|
|
|
else
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
case Stmt::CXXForRangeStmtClass:
|
|
|
|
if (cast<CXXForRangeStmt>(Parent)->getBody() == S)
|
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
|
|
|
break;
|
2009-03-28 11:37:59 +08:00
|
|
|
case Stmt::DoStmtClass:
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-03-28 11:37:59 +08:00
|
|
|
case Stmt::ForStmtClass:
|
|
|
|
if (cast<ForStmt>(Parent)->getBody() == S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-09-09 23:08:12 +08:00
|
|
|
break;
|
2009-03-28 11:37:59 +08:00
|
|
|
case Stmt::IfStmtClass:
|
|
|
|
if (cast<IfStmt>(Parent)->getCond() != S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-04-28 12:23:15 +08:00
|
|
|
break;
|
2009-03-28 11:37:59 +08:00
|
|
|
case Stmt::ObjCForCollectionStmtClass:
|
|
|
|
if (cast<ObjCForCollectionStmt>(Parent)->getBody() == S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-03-28 11:37:59 +08:00
|
|
|
break;
|
|
|
|
case Stmt::WhileStmtClass:
|
|
|
|
if (cast<WhileStmt>(Parent)->getCond() != S)
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-03-28 11:37:59 +08:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2009-03-28 05:16:25 +08:00
|
|
|
S = Parent;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-03-28 05:16:25 +08:00
|
|
|
assert(S && "Cannot have null Stmt for PathDiagnosticLocation");
|
2009-05-12 06:19:32 +08:00
|
|
|
|
2011-09-15 09:08:34 +08:00
|
|
|
return PathDiagnosticLocation(S, SMgr, LC);
|
2009-03-28 05:16:25 +08:00
|
|
|
}
|
|
|
|
|
2009-02-05 07:49:09 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
2009-04-01 07:00:32 +08:00
|
|
|
// "Minimal" path diagnostic generation algorithm.
|
2009-02-05 07:49:09 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
2012-03-17 07:24:20 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
/// If the piece contains a special message, add it to all the call pieces on
|
2019-09-12 04:54:17 +08:00
|
|
|
/// the active stack. For example, my_malloc allocated memory, so MallocChecker
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
/// will construct an event at the call to malloc(), and add a stack hint that
|
|
|
|
/// an allocated memory was returned. We'll use this hint to construct a message
|
|
|
|
/// when returning from the call to my_malloc
|
|
|
|
///
|
|
|
|
/// void *my_malloc() { return malloc(sizeof(int)); }
|
|
|
|
/// void fishy() {
|
|
|
|
/// void *ptr = my_malloc(); // returned allocated memory
|
|
|
|
/// } // leak
|
2019-09-12 04:54:17 +08:00
|
|
|
void PathDiagnosticBuilder::updateStackPiecesWithMessage(
|
|
|
|
PathDiagnosticPieceRef P, const CallWithEntryStack &CallStack) const {
|
|
|
|
if (R->hasCallStackHint(P))
|
|
|
|
for (const auto &I : CallStack) {
|
|
|
|
PathDiagnosticCallPiece *CP = I.first;
|
|
|
|
const ExplodedNode *N = I.second;
|
|
|
|
std::string stackMsg = R->getCallStackMessage(P, N);
|
|
|
|
|
|
|
|
// The last message on the path to final bug is the most important
|
|
|
|
// one. Since we traverse the path backwards, do not add the message
|
|
|
|
// if one has been previously added.
|
|
|
|
if (!CP->hasCallStackMessage())
|
|
|
|
CP->setCallStackMessage(stackMsg);
|
|
|
|
}
|
2012-03-16 05:13:02 +08:00
|
|
|
}
|
2009-02-05 07:49:09 +08:00
|
|
|
|
2018-10-31 22:54:27 +08:00
|
|
|
static void CompactMacroExpandedPieces(PathPieces &path,
|
|
|
|
const SourceManager& SM);
|
2009-04-07 07:06:54 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticPieceRef PathDiagnosticBuilder::generateDiagForSwitchOP(
|
2019-08-14 04:42:48 +08:00
|
|
|
const PathDiagnosticConstruct &C, const CFGBlock *Dst,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation &Start) const {
|
2012-03-16 05:13:02 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const SourceManager &SM = getSourceManager();
|
2018-06-13 03:08:00 +08:00
|
|
|
// Figure out what case arm we took.
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
|
|
|
PathDiagnosticLocation End;
|
2012-03-16 05:13:02 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
if (const Stmt *S = Dst->getLabel()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = PathDiagnosticLocation(S, SM, C.getCurrLocationContext());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
switch (S->getStmtClass()) {
|
|
|
|
default:
|
|
|
|
os << "No cases match in the switch statement. "
|
|
|
|
"Control jumps to line "
|
|
|
|
<< End.asLocation().getExpansionLineNumber();
|
|
|
|
break;
|
|
|
|
case Stmt::DefaultStmtClass:
|
|
|
|
os << "Control jumps to the 'default' case at line "
|
|
|
|
<< End.asLocation().getExpansionLineNumber();
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
2012-07-27 04:04:05 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
case Stmt::CaseStmtClass: {
|
|
|
|
os << "Control jumps to 'case ";
|
|
|
|
const auto *Case = cast<CaseStmt>(S);
|
|
|
|
const Expr *LHS = Case->getLHS()->IgnoreParenCasts();
|
2012-08-30 05:22:37 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
// Determine if it is an enum.
|
|
|
|
bool GetRawInt = true;
|
|
|
|
|
|
|
|
if (const auto *DR = dyn_cast<DeclRefExpr>(LHS)) {
|
|
|
|
// FIXME: Maybe this should be an assertion. Are there cases
|
|
|
|
// were it is not an EnumConstantDecl?
|
|
|
|
const auto *D = dyn_cast<EnumConstantDecl>(DR->getDecl());
|
|
|
|
|
|
|
|
if (D) {
|
|
|
|
GetRawInt = false;
|
|
|
|
os << *D;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (GetRawInt)
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
os << LHS->EvaluateKnownConstInt(getASTContext());
|
2018-06-13 03:08:00 +08:00
|
|
|
|
|
|
|
os << ":' at line " << End.asLocation().getExpansionLineNumber();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
os << "'Default' branch taken. ";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = ExecutionContinues(os, C);
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
2018-06-13 03:08:00 +08:00
|
|
|
return std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
|
|
|
|
os.str());
|
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticPieceRef PathDiagnosticBuilder::generateDiagForGotoOP(
|
2019-08-14 04:42:48 +08:00
|
|
|
const PathDiagnosticConstruct &C, const Stmt *S,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation &Start) const {
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
|
|
|
const PathDiagnosticLocation &End =
|
|
|
|
getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
|
|
|
os << "Control jumps to line " << End.asLocation().getExpansionLineNumber();
|
|
|
|
return std::make_shared<PathDiagnosticControlFlowPiece>(Start, End, os.str());
|
2018-06-13 03:08:00 +08:00
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticPieceRef PathDiagnosticBuilder::generateDiagForBinaryOP(
|
2019-08-14 04:42:48 +08:00
|
|
|
const PathDiagnosticConstruct &C, const Stmt *T, const CFGBlock *Src,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const CFGBlock *Dst) const {
|
|
|
|
|
|
|
|
const SourceManager &SM = getSourceManager();
|
2019-08-14 02:48:08 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
const auto *B = cast<BinaryOperator>(T);
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
|
|
|
os << "Left side of '";
|
|
|
|
PathDiagnosticLocation Start, End;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
if (B->getOpcode() == BO_LAnd) {
|
|
|
|
os << "&&"
|
|
|
|
<< "' is ";
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
if (*(Src->succ_begin() + 1) == Dst) {
|
|
|
|
os << "false";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = PathDiagnosticLocation(B->getLHS(), SM, C.getCurrLocationContext());
|
2018-06-13 03:08:00 +08:00
|
|
|
Start =
|
|
|
|
PathDiagnosticLocation::createOperatorLoc(B, SM);
|
|
|
|
} else {
|
|
|
|
os << "true";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Start =
|
|
|
|
PathDiagnosticLocation(B->getLHS(), SM, C.getCurrLocationContext());
|
|
|
|
End = ExecutionContinues(C);
|
2018-06-13 03:08:00 +08:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
assert(B->getOpcode() == BO_LOr);
|
|
|
|
os << "||"
|
|
|
|
<< "' is ";
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
if (*(Src->succ_begin() + 1) == Dst) {
|
|
|
|
os << "false";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Start =
|
|
|
|
PathDiagnosticLocation(B->getLHS(), SM, C.getCurrLocationContext());
|
|
|
|
End = ExecutionContinues(C);
|
2018-06-13 03:08:00 +08:00
|
|
|
} else {
|
|
|
|
os << "true";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = PathDiagnosticLocation(B->getLHS(), SM, C.getCurrLocationContext());
|
2018-06-13 03:08:00 +08:00
|
|
|
Start =
|
|
|
|
PathDiagnosticLocation::createOperatorLoc(B, SM);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
|
|
|
|
os.str());
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
void PathDiagnosticBuilder::generateMinimalDiagForBlockEdge(
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticConstruct &C, BlockEdge BE) const {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const SourceManager &SM = getSourceManager();
|
|
|
|
const LocationContext *LC = C.getCurrLocationContext();
|
2018-06-13 03:08:00 +08:00
|
|
|
const CFGBlock *Src = BE.getSrc();
|
|
|
|
const CFGBlock *Dst = BE.getDst();
|
2019-05-24 09:34:22 +08:00
|
|
|
const Stmt *T = Src->getTerminatorStmt();
|
2018-06-13 03:08:00 +08:00
|
|
|
if (!T)
|
|
|
|
return;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
auto Start = PathDiagnosticLocation::createBegin(T, SM, LC);
|
|
|
|
switch (T->getStmtClass()) {
|
|
|
|
default:
|
|
|
|
break;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
case Stmt::GotoStmtClass:
|
|
|
|
case Stmt::IndirectGotoStmtClass: {
|
2019-09-12 04:54:21 +08:00
|
|
|
if (const Stmt *S = C.getCurrentNode()->getNextStmtForDiagnostics())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(generateDiagForGotoOP(C, S, Start));
|
2018-06-13 03:08:00 +08:00
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-06-13 03:08:00 +08:00
|
|
|
case Stmt::SwitchStmtClass: {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(generateDiagForSwitchOP(C, Dst, Start));
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
case Stmt::BreakStmtClass:
|
|
|
|
case Stmt::ContinueStmtClass: {
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(os, C);
|
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(Start, End, os.str()));
|
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Determine control-flow for ternary '?'.
|
|
|
|
case Stmt::BinaryConditionalOperatorClass:
|
|
|
|
case Stmt::ConditionalOperatorClass: {
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
|
|
|
os << "'?' condition is ";
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (*(Src->succ_begin() + 1) == Dst)
|
|
|
|
os << "false";
|
|
|
|
else
|
|
|
|
os << "true";
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(C);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(Start, End, os.str()));
|
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Determine control-flow for short-circuited '&&' and '||'.
|
|
|
|
case Stmt::BinaryOperatorClass: {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!C.supportsLogicalOpControlFlow())
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(generateDiagForBinaryOP(C, T, Src, Dst));
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
case Stmt::DoStmtClass:
|
2018-02-08 03:56:52 +08:00
|
|
|
if (*(Src->succ_begin()) == Dst) {
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
os << "Loop condition is true. ";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(os, C);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
|
|
|
|
os.str()));
|
|
|
|
} else {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(C);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(
|
|
|
|
Start, End, "Loop condition is false. Exiting loop"));
|
|
|
|
}
|
|
|
|
break;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
case Stmt::WhileStmtClass:
|
2018-03-07 08:17:48 +08:00
|
|
|
case Stmt::ForStmtClass:
|
2018-02-08 03:56:52 +08:00
|
|
|
if (*(Src->succ_begin() + 1) == Dst) {
|
|
|
|
std::string sbuf;
|
|
|
|
llvm::raw_string_ostream os(sbuf);
|
|
|
|
|
|
|
|
os << "Loop condition is false. ";
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(os, C);
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(Start, End,
|
|
|
|
os.str()));
|
|
|
|
} else {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(C);
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(
|
|
|
|
Start, End, "Loop condition is true. Entering loop body"));
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
case Stmt::IfStmtClass: {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation End = ExecutionContinues(C);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (const Stmt *S = End.asStmt())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
End = getEnclosingStmtLocation(S, C.getCurrLocationContext());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
if (*(Src->succ_begin() + 1) == Dst)
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(
|
|
|
|
Start, End, "Taking false branch"));
|
|
|
|
else
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.getActivePath().push_front(
|
2018-02-08 03:56:52 +08:00
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(
|
|
|
|
Start, End, "Taking true branch"));
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-02-22 13:45:33 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Functions for determining if a loop was executed 0 times.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2013-05-07 09:18:10 +08:00
|
|
|
static bool isLoop(const Stmt *Term) {
|
2013-02-09 03:51:43 +08:00
|
|
|
switch (Term->getStmtClass()) {
|
|
|
|
case Stmt::ForStmtClass:
|
|
|
|
case Stmt::WhileStmtClass:
|
2013-03-14 04:03:31 +08:00
|
|
|
case Stmt::ObjCForCollectionStmtClass:
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
case Stmt::CXXForRangeStmtClass:
|
2013-05-07 09:18:10 +08:00
|
|
|
return true;
|
2013-02-09 03:51:43 +08:00
|
|
|
default:
|
|
|
|
// Note that we intentionally do not include do..while here.
|
|
|
|
return false;
|
|
|
|
}
|
2013-05-07 09:18:10 +08:00
|
|
|
}
|
2013-02-09 03:51:43 +08:00
|
|
|
|
2013-05-07 09:18:10 +08:00
|
|
|
static bool isJumpToFalseBranch(const BlockEdge *BE) {
|
2013-02-09 03:51:43 +08:00
|
|
|
const CFGBlock *Src = BE->getSrc();
|
|
|
|
assert(Src->succ_size() == 2);
|
|
|
|
return (*(Src->succ_begin()+1) == BE->getDst());
|
|
|
|
}
|
|
|
|
|
2019-08-14 02:48:08 +08:00
|
|
|
static bool isContainedByStmt(const ParentMap &PM, const Stmt *S,
|
|
|
|
const Stmt *SubS) {
|
2013-02-22 13:45:33 +08:00
|
|
|
while (SubS) {
|
|
|
|
if (SubS == S)
|
|
|
|
return true;
|
|
|
|
SubS = PM.getParent(SubS);
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2019-08-14 02:48:08 +08:00
|
|
|
static const Stmt *getStmtBeforeCond(const ParentMap &PM, const Stmt *Term,
|
2013-02-22 13:45:33 +08:00
|
|
|
const ExplodedNode *N) {
|
|
|
|
while (N) {
|
|
|
|
Optional<StmtPoint> SP = N->getLocation().getAs<StmtPoint>();
|
|
|
|
if (SP) {
|
|
|
|
const Stmt *S = SP->getStmt();
|
|
|
|
if (!isContainedByStmt(PM, Term, S))
|
|
|
|
return S;
|
|
|
|
}
|
2013-04-24 07:57:43 +08:00
|
|
|
N = N->getFirstPred();
|
2013-02-22 13:45:33 +08:00
|
|
|
}
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
2013-02-22 13:45:33 +08:00
|
|
|
}
|
|
|
|
|
2019-08-14 02:48:08 +08:00
|
|
|
static bool isInLoopBody(const ParentMap &PM, const Stmt *S, const Stmt *Term) {
|
2014-05-27 10:45:47 +08:00
|
|
|
const Stmt *LoopBody = nullptr;
|
2013-02-22 13:45:33 +08:00
|
|
|
switch (Term->getStmtClass()) {
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
case Stmt::CXXForRangeStmtClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *FR = cast<CXXForRangeStmt>(Term);
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
if (isContainedByStmt(PM, FR->getInc(), S))
|
|
|
|
return true;
|
|
|
|
if (isContainedByStmt(PM, FR->getLoopVarStmt(), S))
|
|
|
|
return true;
|
|
|
|
LoopBody = FR->getBody();
|
|
|
|
break;
|
|
|
|
}
|
2013-02-22 13:45:33 +08:00
|
|
|
case Stmt::ForStmtClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *FS = cast<ForStmt>(Term);
|
2013-02-22 13:45:33 +08:00
|
|
|
if (isContainedByStmt(PM, FS->getInc(), S))
|
|
|
|
return true;
|
|
|
|
LoopBody = FS->getBody();
|
|
|
|
break;
|
|
|
|
}
|
2013-03-14 04:03:31 +08:00
|
|
|
case Stmt::ObjCForCollectionStmtClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *FC = cast<ObjCForCollectionStmt>(Term);
|
2013-03-14 04:03:31 +08:00
|
|
|
LoopBody = FC->getBody();
|
|
|
|
break;
|
|
|
|
}
|
2013-02-22 13:45:33 +08:00
|
|
|
case Stmt::WhileStmtClass:
|
|
|
|
LoopBody = cast<WhileStmt>(Term)->getBody();
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return isContainedByStmt(PM, LoopBody, S);
|
|
|
|
}
|
|
|
|
|
2018-05-09 09:00:01 +08:00
|
|
|
/// Adds a sanitized control-flow diagnostic edge to a path.
|
2013-05-04 02:25:33 +08:00
|
|
|
static void addEdgeToPath(PathPieces &path,
|
|
|
|
PathDiagnosticLocation &PrevLoc,
|
2018-09-29 02:49:41 +08:00
|
|
|
PathDiagnosticLocation NewLoc) {
|
2013-05-04 09:13:01 +08:00
|
|
|
if (!NewLoc.isValid())
|
2013-05-04 02:25:33 +08:00
|
|
|
return;
|
|
|
|
|
2013-05-04 09:13:01 +08:00
|
|
|
SourceLocation NewLocL = NewLoc.asLocation();
|
2013-06-08 08:29:24 +08:00
|
|
|
if (NewLocL.isInvalid())
|
2013-05-04 02:25:33 +08:00
|
|
|
return;
|
|
|
|
|
2013-05-31 05:30:17 +08:00
|
|
|
if (!PrevLoc.isValid() || !PrevLoc.asLocation().isValid()) {
|
2013-05-04 02:25:33 +08:00
|
|
|
PrevLoc = NewLoc;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2013-06-06 08:12:37 +08:00
|
|
|
// Ignore self-edges, which occur when there are multiple nodes at the same
|
|
|
|
// statement.
|
|
|
|
if (NewLoc.asStmt() && NewLoc.asStmt() == PrevLoc.asStmt())
|
2013-05-04 02:25:33 +08:00
|
|
|
return;
|
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
path.push_front(
|
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(NewLoc, PrevLoc));
|
2013-05-04 02:25:33 +08:00
|
|
|
PrevLoc = NewLoc;
|
|
|
|
}
|
|
|
|
|
2013-05-21 08:34:40 +08:00
|
|
|
/// A customized wrapper for CFGBlock::getTerminatorCondition()
|
|
|
|
/// which returns the element for ObjCForCollectionStmts.
|
|
|
|
static const Stmt *getTerminatorCondition(const CFGBlock *B) {
|
|
|
|
const Stmt *S = B->getTerminatorCondition();
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *FS = dyn_cast_or_null<ObjCForCollectionStmt>(S))
|
2013-05-21 08:34:40 +08:00
|
|
|
return FS->getElement();
|
|
|
|
return S;
|
|
|
|
}
|
|
|
|
|
2019-08-24 03:59:23 +08:00
|
|
|
constexpr llvm::StringLiteral StrEnteringLoop = "Entering loop body";
|
|
|
|
constexpr llvm::StringLiteral StrLoopBodyZero = "Loop body executed 0 times";
|
|
|
|
constexpr llvm::StringLiteral StrLoopRangeEmpty =
|
|
|
|
"Loop body skipped when range is empty";
|
|
|
|
constexpr llvm::StringLiteral StrLoopCollectionEmpty =
|
2019-08-14 02:48:08 +08:00
|
|
|
"Loop body skipped when collection is empty";
|
2013-06-01 00:56:54 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
static std::unique_ptr<FilesToLineNumsMap>
|
2019-08-14 02:48:08 +08:00
|
|
|
findExecutedLines(const SourceManager &SM, const ExplodedNode *N);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
void PathDiagnosticBuilder::generatePathDiagnosticsForNode(
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticConstruct &C, PathDiagnosticLocation &PrevLoc) const {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
ProgramPoint P = C.getCurrentNode()->getLocation();
|
|
|
|
const SourceManager &SM = getSourceManager();
|
2013-05-04 02:25:33 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Have we encountered an entrance to a call? It may be
|
|
|
|
// the case that we have not encountered a matching
|
|
|
|
// call exit before this point. This means that the path
|
|
|
|
// terminated within the call itself.
|
2018-06-13 03:08:00 +08:00
|
|
|
if (auto CE = P.getAs<CallEnter>()) {
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (C.shouldAddPathEdges()) {
|
2018-06-13 03:08:00 +08:00
|
|
|
// Add an edge to the start of the function.
|
|
|
|
const StackFrameContext *CalleeLC = CE->getCalleeContext();
|
|
|
|
const Decl *D = CalleeLC->getDecl();
|
|
|
|
// Add the edge only when the callee has body. We jump to the beginning
|
|
|
|
// of the *declaration*, however we expect it to be followed by the
|
|
|
|
// body. This isn't the case for autosynthesized property accessors in
|
|
|
|
// Objective-C. No need for a similar extra check for CallExit points
|
|
|
|
// because the exit edge comes from a statement (i.e. return),
|
|
|
|
// not from declaration.
|
|
|
|
if (D->hasBody())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc,
|
|
|
|
PathDiagnosticLocation::createBegin(D, SM));
|
2018-06-13 03:08:00 +08:00
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
|
|
|
|
// Did we visit an entire call?
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
bool VisitedEntireCall = C.PD->isWithinCall();
|
|
|
|
C.PD->popActivePath();
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticCallPiece *Call;
|
2018-02-08 03:56:52 +08:00
|
|
|
if (VisitedEntireCall) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Call = cast<PathDiagnosticCallPiece>(C.getActivePath().front().get());
|
2018-02-08 03:56:52 +08:00
|
|
|
} else {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
// The path terminated within a nested location context, create a new
|
|
|
|
// call piece to encapsulate the rest of the path pieces.
|
2018-02-08 03:56:52 +08:00
|
|
|
const Decl *Caller = CE->getLocationContext()->getDecl();
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Call = PathDiagnosticCallPiece::construct(C.getActivePath(), Caller);
|
|
|
|
assert(C.getActivePath().size() == 1 &&
|
|
|
|
C.getActivePath().front().get() == Call);
|
|
|
|
|
|
|
|
// Since we just transferred the path over to the call piece, reset the
|
|
|
|
// mapping of the active path to the current location context.
|
|
|
|
assert(C.isInLocCtxMap(&C.getActivePath()) &&
|
|
|
|
"When we ascend to a previously unvisited call, the active path's "
|
|
|
|
"address shouldn't change, but rather should be compacted into "
|
|
|
|
"a single CallEvent!");
|
|
|
|
C.updateLocCtxMap(&C.getActivePath(), C.getCurrLocationContext());
|
|
|
|
|
|
|
|
// Record the location context mapping for the path within the call.
|
|
|
|
assert(!C.isInLocCtxMap(&Call->path) &&
|
|
|
|
"When we ascend to a previously unvisited call, this must be the "
|
|
|
|
"first time we encounter the caller context!");
|
|
|
|
C.updateLocCtxMap(&Call->path, CE->getCalleeContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Call->setCallee(*CE, SM);
|
2013-05-31 05:30:17 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Update the previous location in the active path.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PrevLoc = Call->getLocation();
|
2013-05-08 05:12:06 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!C.CallStack.empty()) {
|
|
|
|
assert(C.CallStack.back().first == Call);
|
|
|
|
C.CallStack.pop_back();
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
2013-05-08 05:12:06 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
assert(C.getCurrLocationContext() == C.getLocationContextForActivePath() &&
|
|
|
|
"The current position in the bug path is out of sync with the "
|
|
|
|
"location context associated with the active path!");
|
2018-02-08 03:56:52 +08:00
|
|
|
|
|
|
|
// Have we encountered an exit from a function call?
|
|
|
|
if (Optional<CallExitEnd> CE = P.getAs<CallExitEnd>()) {
|
2013-05-25 05:43:05 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// We are descending into a call (backwards). Construct
|
|
|
|
// a new call piece to contain the path pieces for that call.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
auto Call = PathDiagnosticCallPiece::construct(*CE, SM);
|
2018-06-13 03:08:00 +08:00
|
|
|
// Record the mapping from call piece to LocationContext.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
assert(!C.isInLocCtxMap(&Call->path) &&
|
|
|
|
"We just entered a call, this must've been the first time we "
|
|
|
|
"encounter its context!");
|
|
|
|
C.updateLocCtxMap(&Call->path, CE->getCalleeContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (C.shouldAddPathEdges()) {
|
2018-06-13 03:08:00 +08:00
|
|
|
// Add the edge to the return site.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc, Call->callReturn);
|
2018-06-13 03:08:00 +08:00
|
|
|
PrevLoc.invalidate();
|
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
auto *P = Call.get();
|
|
|
|
C.getActivePath().push_front(std::move(Call));
|
2018-02-08 03:56:52 +08:00
|
|
|
|
|
|
|
// Make the contents of the call the active path for now.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
C.PD->pushActivePath(&P->path);
|
2019-08-14 04:42:48 +08:00
|
|
|
C.CallStack.push_back(CallWithEntry(P, C.getCurrentNode()));
|
2018-06-13 03:08:00 +08:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (auto PS = P.getAs<PostStmt>()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!C.shouldAddPathEdges())
|
2018-06-13 03:08:00 +08:00
|
|
|
return;
|
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Add an edge. If this is an ObjCForCollectionStmt do
|
|
|
|
// not add an edge here as it appears in the CFG both
|
|
|
|
// as a terminator and as a terminator condition.
|
|
|
|
if (!isa<ObjCForCollectionStmt>(PS->getStmt())) {
|
|
|
|
PathDiagnosticLocation L =
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation(PS->getStmt(), SM, C.getCurrLocationContext());
|
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc, L);
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
2018-06-13 03:08:00 +08:00
|
|
|
|
|
|
|
} else if (auto BE = P.getAs<BlockEdge>()) {
|
|
|
|
|
2020-12-01 16:53:53 +08:00
|
|
|
if (C.shouldAddControlNotes()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
generateMinimalDiagForBlockEdge(C, *BE);
|
2020-12-01 16:53:53 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!C.shouldAddPathEdges()) {
|
2018-06-13 03:08:00 +08:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
// Are we jumping to the head of a loop? Add a special diagnostic.
|
|
|
|
if (const Stmt *Loop = BE->getSrc()->getLoopTarget()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation L(Loop, SM, C.getCurrLocationContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
const Stmt *Body = nullptr;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *FS = dyn_cast<ForStmt>(Loop))
|
2018-02-08 03:56:52 +08:00
|
|
|
Body = FS->getBody();
|
2018-03-07 08:17:48 +08:00
|
|
|
else if (const auto *WS = dyn_cast<WhileStmt>(Loop))
|
2018-02-08 03:56:52 +08:00
|
|
|
Body = WS->getBody();
|
2018-03-07 08:17:48 +08:00
|
|
|
else if (const auto *OFS = dyn_cast<ObjCForCollectionStmt>(Loop)) {
|
2018-02-08 03:56:52 +08:00
|
|
|
Body = OFS->getBody();
|
2018-03-07 08:17:48 +08:00
|
|
|
} else if (const auto *FRS = dyn_cast<CXXForRangeStmt>(Loop)) {
|
2018-02-08 03:56:52 +08:00
|
|
|
Body = FRS->getBody();
|
|
|
|
}
|
|
|
|
// do-while statements are explicitly excluded here
|
2013-05-04 02:25:33 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
auto p = std::make_shared<PathDiagnosticEventPiece>(
|
2020-12-01 16:53:53 +08:00
|
|
|
L, "Looping back to the head of the loop");
|
2018-02-08 03:56:52 +08:00
|
|
|
p->setPrunable(true);
|
2013-05-04 02:25:33 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc, p->getLocation());
|
2020-12-01 16:53:53 +08:00
|
|
|
// We might've added a very similar control node already
|
|
|
|
if (!C.shouldAddControlNotes()) {
|
|
|
|
C.getActivePath().push_front(std::move(p));
|
|
|
|
}
|
2013-05-04 02:25:33 +08:00
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc,
|
|
|
|
PathDiagnosticLocation::createEndBrace(CS, SM));
|
2013-05-04 02:25:33 +08:00
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
2013-05-04 02:25:33 +08:00
|
|
|
|
2018-02-08 03:56:52 +08:00
|
|
|
const CFGBlock *BSrc = BE->getSrc();
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const ParentMap &PM = C.getParentMap();
|
2018-02-08 03:56:52 +08:00
|
|
|
|
2019-05-24 09:34:22 +08:00
|
|
|
if (const Stmt *Term = BSrc->getTerminatorStmt()) {
|
2018-02-08 03:56:52 +08:00
|
|
|
// Are we jumping past the loop body without ever executing the
|
|
|
|
// loop (because the condition was false)?
|
|
|
|
if (isLoop(Term)) {
|
|
|
|
const Stmt *TermCond = getTerminatorCondition(BSrc);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
bool IsInLoopBody = isInLoopBody(
|
|
|
|
PM, getStmtBeforeCond(PM, TermCond, C.getCurrentNode()), Term);
|
2018-02-08 03:56:52 +08:00
|
|
|
|
2019-08-14 02:48:08 +08:00
|
|
|
StringRef str;
|
2018-02-08 03:56:52 +08:00
|
|
|
|
|
|
|
if (isJumpToFalseBranch(&*BE)) {
|
|
|
|
if (!IsInLoopBody) {
|
|
|
|
if (isa<ObjCForCollectionStmt>(Term)) {
|
|
|
|
str = StrLoopCollectionEmpty;
|
|
|
|
} else if (isa<CXXForRangeStmt>(Term)) {
|
|
|
|
str = StrLoopRangeEmpty;
|
|
|
|
} else {
|
|
|
|
str = StrLoopBodyZero;
|
|
|
|
}
|
2013-05-04 02:25:33 +08:00
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
} else {
|
|
|
|
str = StrEnteringLoop;
|
2013-05-04 02:25:33 +08:00
|
|
|
}
|
|
|
|
|
2019-08-14 02:48:08 +08:00
|
|
|
if (!str.empty()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation L(TermCond ? TermCond : Term, SM,
|
|
|
|
C.getCurrLocationContext());
|
2018-02-08 03:56:52 +08:00
|
|
|
auto PE = std::make_shared<PathDiagnosticEventPiece>(L, str);
|
|
|
|
PE->setPrunable(true);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc, PE->getLocation());
|
2020-12-01 16:53:53 +08:00
|
|
|
|
|
|
|
// We might've added a very similar control node already
|
|
|
|
if (!C.shouldAddControlNotes()) {
|
|
|
|
C.getActivePath().push_front(std::move(PE));
|
|
|
|
}
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
2021-10-20 23:43:31 +08:00
|
|
|
} else if (isa<BreakStmt, ContinueStmt, GotoStmt>(Term)) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation L(Term, SM, C.getCurrLocationContext());
|
|
|
|
addEdgeToPath(C.getActivePath(), PrevLoc, L);
|
2018-02-08 03:56:52 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2013-05-04 02:25:33 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
static std::unique_ptr<PathDiagnostic>
|
2019-09-10 04:34:40 +08:00
|
|
|
generateDiagnosticForBasicReport(const BasicBugReport *R) {
|
2019-01-19 03:24:55 +08:00
|
|
|
const BugType &BT = R->getBugType();
|
2019-08-15 07:04:18 +08:00
|
|
|
return std::make_unique<PathDiagnostic>(
|
2019-09-13 03:09:24 +08:00
|
|
|
BT.getCheckerName(), R->getDeclWithIssue(), BT.getDescription(),
|
2019-09-10 04:34:40 +08:00
|
|
|
R->getDescription(), R->getShortDescription(/*UseFallback=*/false),
|
|
|
|
BT.getCategory(), R->getUniqueingLocation(), R->getUniqueingDecl(),
|
|
|
|
std::make_unique<FilesToLineNumsMap>());
|
|
|
|
}
|
|
|
|
|
|
|
|
static std::unique_ptr<PathDiagnostic>
|
|
|
|
generateEmptyDiagnosticForReport(const PathSensitiveBugReport *R,
|
|
|
|
const SourceManager &SM) {
|
|
|
|
const BugType &BT = R->getBugType();
|
|
|
|
return std::make_unique<PathDiagnostic>(
|
2019-09-13 03:09:24 +08:00
|
|
|
BT.getCheckerName(), R->getDeclWithIssue(), BT.getDescription(),
|
2019-09-10 04:34:40 +08:00
|
|
|
R->getDescription(), R->getShortDescription(/*UseFallback=*/false),
|
|
|
|
BT.getCategory(), R->getUniqueingLocation(), R->getUniqueingDecl(),
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
findExecutedLines(SM, R->getErrorNode()));
|
2013-05-04 02:25:33 +08:00
|
|
|
}
|
|
|
|
|
2013-06-04 06:59:48 +08:00
|
|
|
static const Stmt *getStmtParent(const Stmt *S, const ParentMap &PM) {
|
2013-05-04 09:13:01 +08:00
|
|
|
if (!S)
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
2013-05-17 14:48:27 +08:00
|
|
|
|
|
|
|
while (true) {
|
|
|
|
S = PM.getParentIgnoreParens(S);
|
|
|
|
|
|
|
|
if (!S)
|
|
|
|
break;
|
|
|
|
|
2021-10-20 23:43:31 +08:00
|
|
|
if (isa<FullExpr, CXXBindTemporaryExpr, SubstNonTypeTemplateParmExpr>(S))
|
2013-05-17 14:48:27 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return S;
|
2013-05-04 09:13:01 +08:00
|
|
|
}
|
|
|
|
|
2013-05-04 09:13:22 +08:00
|
|
|
static bool isConditionForTerminator(const Stmt *S, const Stmt *Cond) {
|
|
|
|
switch (S->getStmtClass()) {
|
2013-05-08 05:11:57 +08:00
|
|
|
case Stmt::BinaryOperatorClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *BO = cast<BinaryOperator>(S);
|
2013-05-08 05:11:57 +08:00
|
|
|
if (!BO->isLogicalOp())
|
|
|
|
return false;
|
|
|
|
return BO->getLHS() == Cond || BO->getRHS() == Cond;
|
|
|
|
}
|
2013-05-07 15:30:07 +08:00
|
|
|
case Stmt::IfStmtClass:
|
|
|
|
return cast<IfStmt>(S)->getCond() == Cond;
|
2013-05-04 09:13:22 +08:00
|
|
|
case Stmt::ForStmtClass:
|
|
|
|
return cast<ForStmt>(S)->getCond() == Cond;
|
|
|
|
case Stmt::WhileStmtClass:
|
|
|
|
return cast<WhileStmt>(S)->getCond() == Cond;
|
|
|
|
case Stmt::DoStmtClass:
|
|
|
|
return cast<DoStmt>(S)->getCond() == Cond;
|
|
|
|
case Stmt::ChooseExprClass:
|
|
|
|
return cast<ChooseExpr>(S)->getCond() == Cond;
|
|
|
|
case Stmt::IndirectGotoStmtClass:
|
|
|
|
return cast<IndirectGotoStmt>(S)->getTarget() == Cond;
|
|
|
|
case Stmt::SwitchStmtClass:
|
|
|
|
return cast<SwitchStmt>(S)->getCond() == Cond;
|
|
|
|
case Stmt::BinaryConditionalOperatorClass:
|
|
|
|
return cast<BinaryConditionalOperator>(S)->getCond() == Cond;
|
2013-05-08 09:15:20 +08:00
|
|
|
case Stmt::ConditionalOperatorClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *CO = cast<ConditionalOperator>(S);
|
2013-05-08 09:15:20 +08:00
|
|
|
return CO->getCond() == Cond ||
|
|
|
|
CO->getLHS() == Cond ||
|
|
|
|
CO->getRHS() == Cond;
|
|
|
|
}
|
2013-05-04 09:13:22 +08:00
|
|
|
case Stmt::ObjCForCollectionStmtClass:
|
|
|
|
return cast<ObjCForCollectionStmt>(S)->getElement() == Cond;
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
case Stmt::CXXForRangeStmtClass: {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *FRS = cast<CXXForRangeStmt>(S);
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
return FRS->getCond() == Cond || FRS->getRangeInit() == Cond;
|
|
|
|
}
|
2013-05-04 09:13:22 +08:00
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-05-08 05:11:49 +08:00
|
|
|
static bool isIncrementOrInitInForLoop(const Stmt *S, const Stmt *FL) {
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *FS = dyn_cast<ForStmt>(FL))
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
return FS->getInc() == S || FS->getInit() == S;
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *FRS = dyn_cast<CXXForRangeStmt>(FL))
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
return FRS->getInc() == S || FRS->getRangeStmt() == S ||
|
|
|
|
FRS->getLoopVarStmt() || FRS->getRangeInit() == S;
|
|
|
|
return false;
|
2013-05-08 01:02:41 +08:00
|
|
|
}
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
using OptimizedCallsSet = llvm::DenseSet<const PathDiagnosticCallPiece *>;
|
2013-05-04 09:13:22 +08:00
|
|
|
|
2013-06-04 06:59:48 +08:00
|
|
|
/// Adds synthetic edges from top-level statements to their subexpressions.
|
|
|
|
///
|
|
|
|
/// This avoids a "swoosh" effect, where an edge from a top-level statement A
|
|
|
|
/// points to a sub-expression B.1 that's not at the start of B. In these cases,
|
|
|
|
/// we'd like to see an edge from A to B, then another one from B to B.1.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
static void addContextEdges(PathPieces &pieces, const LocationContext *LC) {
|
|
|
|
const ParentMap &PM = LC->getParentMap();
|
2013-06-04 06:59:48 +08:00
|
|
|
PathPieces::iterator Prev = pieces.end();
|
|
|
|
for (PathPieces::iterator I = pieces.begin(), E = Prev; I != E;
|
|
|
|
Prev = I, ++I) {
|
2018-03-07 08:17:48 +08:00
|
|
|
auto *Piece = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
if (!Piece)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
PathDiagnosticLocation SrcLoc = Piece->getStartLocation();
|
2013-06-04 06:59:53 +08:00
|
|
|
SmallVector<PathDiagnosticLocation, 4> SrcContexts;
|
|
|
|
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
PathDiagnosticLocation NextSrcContext = SrcLoc;
|
2014-05-27 10:45:47 +08:00
|
|
|
const Stmt *InnerStmt = nullptr;
|
2013-06-04 06:59:53 +08:00
|
|
|
while (NextSrcContext.isValid() && NextSrcContext.asStmt() != InnerStmt) {
|
|
|
|
SrcContexts.push_back(NextSrcContext);
|
|
|
|
InnerStmt = NextSrcContext.asStmt();
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
NextSrcContext = getEnclosingStmtLocation(InnerStmt, LC,
|
[analyzer; new edges] Simplify edges in a C++11 for-range loop.
Previously our edges were completely broken here; now, the final result
is a very simple set of edges in most cases: one up to the "for" keyword
for context, and one into the body of the loop. This matches the behavior
for ObjC for-in loops.
In the AST, however, CXXForRangeStmts are handled very differently from
ObjCForCollectionStmts. Since they are specified in terms of equivalent
statements in the C++ standard, we actually have implicit AST nodes for
all of the semantic statements. This makes evaluation very easy, but
diagnostic locations a bit trickier. Fortunately, the problem can be
generally defined away by marking all of the implicit statements as
part of the top-level for-range statement.
One of the implicit statements in a for-range statement is the declaration
of implicit iterators __begin and __end. The CFG synthesizes two
separate DeclStmts to match each of these decls, but until now these
synthetic DeclStmts weren't in the function's ParentMap. Now, the CFG
keeps track of its synthetic statements, and the AnalysisDeclContext will
make sure to add them to the ParentMap.
<rdar://problem/14038483>
llvm-svn: 183449
2013-06-07 05:53:45 +08:00
|
|
|
/*allowNested=*/true);
|
2013-06-04 06:59:53 +08:00
|
|
|
}
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
// Repeatedly split the edge as necessary.
|
|
|
|
// This is important for nested logical expressions (||, &&, ?:) where we
|
|
|
|
// want to show all the levels of context.
|
|
|
|
while (true) {
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *Dst = Piece->getEndLocation().getStmtOrNull();
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
// We are looking at an edge. Is the destination within a larger
|
|
|
|
// expression?
|
|
|
|
PathDiagnosticLocation DstContext =
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
getEnclosingStmtLocation(Dst, LC, /*allowNested=*/true);
|
2013-06-04 06:59:48 +08:00
|
|
|
if (!DstContext.isValid() || DstContext.asStmt() == Dst)
|
|
|
|
break;
|
|
|
|
|
|
|
|
// If the source is in the same context, we're already good.
|
2021-10-14 11:41:54 +08:00
|
|
|
if (llvm::is_contained(SrcContexts, DstContext))
|
2013-06-04 06:59:48 +08:00
|
|
|
break;
|
|
|
|
|
|
|
|
// Update the subexpression node to point to the context edge.
|
|
|
|
Piece->setStartLocation(DstContext);
|
|
|
|
|
|
|
|
// Try to extend the previous edge if it's at the same level as the source
|
|
|
|
// context.
|
|
|
|
if (Prev != E) {
|
2017-01-06 01:26:53 +08:00
|
|
|
auto *PrevPiece = dyn_cast<PathDiagnosticControlFlowPiece>(Prev->get());
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
if (PrevPiece) {
|
2018-06-13 04:51:01 +08:00
|
|
|
if (const Stmt *PrevSrc =
|
|
|
|
PrevPiece->getStartLocation().getStmtOrNull()) {
|
2013-06-04 06:59:48 +08:00
|
|
|
const Stmt *PrevSrcParent = getStmtParent(PrevSrc, PM);
|
2018-06-13 04:51:01 +08:00
|
|
|
if (PrevSrcParent ==
|
|
|
|
getStmtParent(DstContext.getStmtOrNull(), PM)) {
|
2013-06-04 06:59:48 +08:00
|
|
|
PrevPiece->setEndLocation(DstContext);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Otherwise, split the current edge into a context edge and a
|
|
|
|
// subexpression edge. Note that the context statement may itself have
|
|
|
|
// context.
|
2017-01-06 01:26:53 +08:00
|
|
|
auto P =
|
|
|
|
std::make_shared<PathDiagnosticControlFlowPiece>(SrcLoc, DstContext);
|
|
|
|
Piece = P.get();
|
|
|
|
I = pieces.insert(I, std::move(P));
|
2013-06-04 06:59:48 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-09 09:00:01 +08:00
|
|
|
/// Move edges from a branch condition to a branch target
|
2013-06-04 06:59:48 +08:00
|
|
|
/// when the condition is simple.
|
|
|
|
///
|
|
|
|
/// This restructures some of the work of addContextEdges. That function
|
|
|
|
/// creates edges this may destroy, but they work together to create a more
|
|
|
|
/// aesthetically set of edges around branches. After the call to
|
|
|
|
/// addContextEdges, we may have (1) an edge to the branch, (2) an edge from
|
|
|
|
/// the branch to the branch condition, and (3) an edge from the branch
|
|
|
|
/// condition to the branch target. We keep (1), but may wish to remove (2)
|
|
|
|
/// and move the source of (3) to the branch if the branch condition is simple.
|
|
|
|
static void simplifySimpleBranches(PathPieces &pieces) {
|
|
|
|
for (PathPieces::iterator I = pieces.begin(), E = pieces.end(); I != E; ++I) {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
if (!PieceI)
|
|
|
|
continue;
|
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s1Start = PieceI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s1End = PieceI->getEndLocation().getStmtOrNull();
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
if (!s1Start || !s1End)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
PathPieces::iterator NextI = I; ++NextI;
|
|
|
|
if (NextI == E)
|
|
|
|
break;
|
|
|
|
|
2014-05-27 10:45:47 +08:00
|
|
|
PathDiagnosticControlFlowPiece *PieceNextI = nullptr;
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
while (true) {
|
|
|
|
if (NextI == E)
|
|
|
|
break;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *EV = dyn_cast<PathDiagnosticEventPiece>(NextI->get());
|
2013-06-04 06:59:48 +08:00
|
|
|
if (EV) {
|
|
|
|
StringRef S = EV->getString();
|
2013-11-08 09:15:30 +08:00
|
|
|
if (S == StrEnteringLoop || S == StrLoopBodyZero ||
|
|
|
|
S == StrLoopCollectionEmpty || S == StrLoopRangeEmpty) {
|
2013-06-04 06:59:48 +08:00
|
|
|
++NextI;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2017-01-06 01:26:53 +08:00
|
|
|
PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
|
2013-06-04 06:59:48 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!PieceNextI)
|
|
|
|
continue;
|
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s2Start = PieceNextI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s2End = PieceNextI->getEndLocation().getStmtOrNull();
|
2013-06-04 06:59:48 +08:00
|
|
|
|
|
|
|
if (!s2Start || !s2End || s1End != s2Start)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// We only perform this transformation for specific branch kinds.
|
|
|
|
// We don't want to do this for do..while, for example.
|
2021-10-20 23:43:31 +08:00
|
|
|
if (!isa<ForStmt, WhileStmt, IfStmt, ObjCForCollectionStmt,
|
|
|
|
CXXForRangeStmt>(s1Start))
|
2013-06-04 06:59:48 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// Is s1End the branch condition?
|
|
|
|
if (!isConditionForTerminator(s1Start, s1End))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Perform the hoisting by eliminating (2) and changing the start
|
|
|
|
// location of (3).
|
|
|
|
PieceNextI->setStartLocation(PieceI->getStartLocation());
|
|
|
|
I = pieces.erase(I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-06-04 07:00:09 +08:00
|
|
|
/// Returns the number of bytes in the given (character-based) SourceRange.
|
2013-06-04 07:00:05 +08:00
|
|
|
///
|
|
|
|
/// If the locations in the range are not on the same line, returns None.
|
|
|
|
///
|
|
|
|
/// Note that this does not do a precise user-visible character or column count.
|
2019-08-14 02:48:08 +08:00
|
|
|
static Optional<size_t> getLengthOnSingleLine(const SourceManager &SM,
|
2013-06-04 07:00:05 +08:00
|
|
|
SourceRange Range) {
|
|
|
|
SourceRange ExpansionRange(SM.getExpansionLoc(Range.getBegin()),
|
2018-04-30 13:25:48 +08:00
|
|
|
SM.getExpansionRange(Range.getEnd()).getEnd());
|
2013-06-04 07:00:05 +08:00
|
|
|
|
|
|
|
FileID FID = SM.getFileID(ExpansionRange.getBegin());
|
|
|
|
if (FID != SM.getFileID(ExpansionRange.getEnd()))
|
|
|
|
return None;
|
|
|
|
|
2020-10-15 02:06:37 +08:00
|
|
|
Optional<MemoryBufferRef> Buffer = SM.getBufferOrNone(FID);
|
|
|
|
if (!Buffer)
|
2013-06-04 07:00:05 +08:00
|
|
|
return None;
|
|
|
|
|
|
|
|
unsigned BeginOffset = SM.getFileOffset(ExpansionRange.getBegin());
|
|
|
|
unsigned EndOffset = SM.getFileOffset(ExpansionRange.getEnd());
|
|
|
|
StringRef Snippet = Buffer->getBuffer().slice(BeginOffset, EndOffset);
|
|
|
|
|
|
|
|
// We're searching the raw bytes of the buffer here, which might include
|
|
|
|
// escaped newlines and such. That's okay; we're trying to decide whether the
|
|
|
|
// SourceRange is covering a large or small amount of space in the user's
|
|
|
|
// editor.
|
|
|
|
if (Snippet.find_first_of("\r\n") != StringRef::npos)
|
|
|
|
return None;
|
|
|
|
|
|
|
|
// This isn't Unicode-aware, but it doesn't need to be.
|
|
|
|
return Snippet.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// \sa getLengthOnSingleLine(SourceManager, SourceRange)
|
2019-08-14 02:48:08 +08:00
|
|
|
static Optional<size_t> getLengthOnSingleLine(const SourceManager &SM,
|
2013-06-04 07:00:05 +08:00
|
|
|
const Stmt *S) {
|
|
|
|
return getLengthOnSingleLine(SM, S->getSourceRange());
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Eliminate two-edge cycles created by addContextEdges().
|
|
|
|
///
|
|
|
|
/// Once all the context edges are in place, there are plenty of cases where
|
|
|
|
/// there's a single edge from a top-level statement to a subexpression,
|
|
|
|
/// followed by a single path note, and then a reverse edge to get back out to
|
|
|
|
/// the top level. If the statement is simple enough, the subexpression edges
|
|
|
|
/// just add noise and make it harder to understand what's going on.
|
|
|
|
///
|
|
|
|
/// This function only removes edges in pairs, because removing only one edge
|
|
|
|
/// might leave other edges dangling.
|
|
|
|
///
|
|
|
|
/// This will not remove edges in more complicated situations:
|
|
|
|
/// - if there is more than one "hop" leading to or from a subexpression.
|
|
|
|
/// - if there is an inlined call between the edges instead of a single event.
|
|
|
|
/// - if the whole statement is large enough that having subexpression arrows
|
|
|
|
/// might be helpful.
|
2019-08-14 02:48:08 +08:00
|
|
|
static void removeContextCycles(PathPieces &Path, const SourceManager &SM) {
|
2013-06-04 07:00:00 +08:00
|
|
|
for (PathPieces::iterator I = Path.begin(), E = Path.end(); I != E; ) {
|
|
|
|
// Pattern match the current piece and its successor.
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
|
2013-06-04 07:00:00 +08:00
|
|
|
|
|
|
|
if (!PieceI) {
|
|
|
|
++I;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s1Start = PieceI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s1End = PieceI->getEndLocation().getStmtOrNull();
|
2013-06-04 07:00:00 +08:00
|
|
|
|
|
|
|
PathPieces::iterator NextI = I; ++NextI;
|
|
|
|
if (NextI == E)
|
|
|
|
break;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceNextI =
|
2017-01-06 01:26:53 +08:00
|
|
|
dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
|
2013-06-04 07:00:00 +08:00
|
|
|
|
|
|
|
if (!PieceNextI) {
|
2017-01-06 01:26:53 +08:00
|
|
|
if (isa<PathDiagnosticEventPiece>(NextI->get())) {
|
2013-06-04 07:00:00 +08:00
|
|
|
++NextI;
|
|
|
|
if (NextI == E)
|
|
|
|
break;
|
2017-01-06 01:26:53 +08:00
|
|
|
PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
|
2013-06-04 07:00:00 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!PieceNextI) {
|
|
|
|
++I;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s2Start = PieceNextI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s2End = PieceNextI->getEndLocation().getStmtOrNull();
|
2013-06-04 07:00:00 +08:00
|
|
|
|
|
|
|
if (s1Start && s2Start && s1Start == s2End && s2Start == s1End) {
|
2013-06-04 07:00:05 +08:00
|
|
|
const size_t MAX_SHORT_LINE_LENGTH = 80;
|
|
|
|
Optional<size_t> s1Length = getLengthOnSingleLine(SM, s1Start);
|
|
|
|
if (s1Length && *s1Length <= MAX_SHORT_LINE_LENGTH) {
|
|
|
|
Optional<size_t> s2Length = getLengthOnSingleLine(SM, s2Start);
|
|
|
|
if (s2Length && *s2Length <= MAX_SHORT_LINE_LENGTH) {
|
|
|
|
Path.erase(I);
|
|
|
|
I = Path.erase(NextI);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2013-06-04 07:00:00 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
++I;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-05-09 09:00:01 +08:00
|
|
|
/// Return true if X is contained by Y.
|
2019-08-14 02:48:08 +08:00
|
|
|
static bool lexicalContains(const ParentMap &PM, const Stmt *X, const Stmt *Y) {
|
2013-05-22 05:38:05 +08:00
|
|
|
while (X) {
|
|
|
|
if (X == Y)
|
|
|
|
return true;
|
|
|
|
X = PM.getParent(X);
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2013-05-23 02:52:35 +08:00
|
|
|
// Remove short edges on the same line less than 3 columns in difference.
|
2019-08-14 02:48:08 +08:00
|
|
|
static void removePunyEdges(PathPieces &path, const SourceManager &SM,
|
|
|
|
const ParentMap &PM) {
|
2013-05-23 02:52:35 +08:00
|
|
|
bool erased = false;
|
|
|
|
|
|
|
|
for (PathPieces::iterator I = path.begin(), E = path.end(); I != E;
|
|
|
|
erased ? I : ++I) {
|
|
|
|
erased = false;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
|
2013-05-23 02:52:35 +08:00
|
|
|
|
|
|
|
if (!PieceI)
|
|
|
|
continue;
|
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *start = PieceI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *end = PieceI->getEndLocation().getStmtOrNull();
|
2013-05-23 02:52:35 +08:00
|
|
|
|
|
|
|
if (!start || !end)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
const Stmt *endParent = PM.getParent(end);
|
|
|
|
if (!endParent)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (isConditionForTerminator(end, endParent))
|
|
|
|
continue;
|
|
|
|
|
2018-08-10 05:08:08 +08:00
|
|
|
SourceLocation FirstLoc = start->getBeginLoc();
|
|
|
|
SourceLocation SecondLoc = end->getBeginLoc();
|
2013-05-23 02:52:35 +08:00
|
|
|
|
2013-08-22 08:27:10 +08:00
|
|
|
if (!SM.isWrittenInSameFile(FirstLoc, SecondLoc))
|
2013-05-23 02:52:35 +08:00
|
|
|
continue;
|
2013-06-04 07:00:05 +08:00
|
|
|
if (SM.isBeforeInTranslationUnit(SecondLoc, FirstLoc))
|
|
|
|
std::swap(SecondLoc, FirstLoc);
|
2013-05-23 02:52:35 +08:00
|
|
|
|
2013-06-04 07:00:05 +08:00
|
|
|
SourceRange EdgeRange(FirstLoc, SecondLoc);
|
|
|
|
Optional<size_t> ByteWidth = getLengthOnSingleLine(SM, EdgeRange);
|
2013-05-23 02:52:35 +08:00
|
|
|
|
2013-06-04 07:00:05 +08:00
|
|
|
// If the statements are on different lines, continue.
|
|
|
|
if (!ByteWidth)
|
2013-05-23 02:52:35 +08:00
|
|
|
continue;
|
|
|
|
|
2013-06-04 07:00:05 +08:00
|
|
|
const size_t MAX_PUNY_EDGE_LENGTH = 2;
|
|
|
|
if (*ByteWidth <= MAX_PUNY_EDGE_LENGTH) {
|
|
|
|
// FIXME: There are enough /bytes/ between the endpoints of the edge, but
|
|
|
|
// there might not be enough /columns/. A proper user-visible column count
|
|
|
|
// is probably too expensive, though.
|
2013-05-23 03:25:03 +08:00
|
|
|
I = path.erase(I);
|
2013-05-23 02:52:35 +08:00
|
|
|
erased = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-05-23 03:10:41 +08:00
|
|
|
static void removeIdenticalEvents(PathPieces &path) {
|
|
|
|
for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ++I) {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceI = dyn_cast<PathDiagnosticEventPiece>(I->get());
|
2013-05-23 03:10:41 +08:00
|
|
|
|
|
|
|
if (!PieceI)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
PathPieces::iterator NextI = I; ++NextI;
|
|
|
|
if (NextI == E)
|
|
|
|
return;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceNextI = dyn_cast<PathDiagnosticEventPiece>(NextI->get());
|
2013-05-23 03:10:41 +08:00
|
|
|
|
|
|
|
if (!PieceNextI)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Erase the second piece if it has the same exact message text.
|
|
|
|
if (PieceI->getString() == PieceNextI->getString()) {
|
|
|
|
path.erase(NextI);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
static bool optimizeEdges(const PathDiagnosticConstruct &C, PathPieces &path,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
OptimizedCallsSet &OCS) {
|
2013-05-04 09:13:01 +08:00
|
|
|
bool hasChanges = false;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const LocationContext *LC = C.getLocationContextFor(&path);
|
2013-05-04 09:13:01 +08:00
|
|
|
assert(LC);
|
2019-08-14 02:48:08 +08:00
|
|
|
const ParentMap &PM = LC->getParentMap();
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const SourceManager &SM = C.getSourceManager();
|
2013-05-04 09:13:01 +08:00
|
|
|
|
2013-05-07 05:59:37 +08:00
|
|
|
for (PathPieces::iterator I = path.begin(), E = path.end(); I != E; ) {
|
2013-05-04 09:13:01 +08:00
|
|
|
// Optimize subpaths.
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *CallI = dyn_cast<PathDiagnosticCallPiece>(I->get())) {
|
2013-05-07 05:59:37 +08:00
|
|
|
// Record the fact that a call has been optimized so we only do the
|
|
|
|
// effort once.
|
2013-05-04 09:13:22 +08:00
|
|
|
if (!OCS.count(CallI)) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
while (optimizeEdges(C, CallI->path, OCS)) {
|
|
|
|
}
|
2013-05-04 09:13:22 +08:00
|
|
|
OCS.insert(CallI);
|
|
|
|
}
|
2013-05-07 05:59:37 +08:00
|
|
|
++I;
|
2013-05-04 09:13:01 +08:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Pattern match the current piece and its successor.
|
2017-01-06 01:26:53 +08:00
|
|
|
auto *PieceI = dyn_cast<PathDiagnosticControlFlowPiece>(I->get());
|
2013-05-04 09:13:01 +08:00
|
|
|
|
2013-05-07 05:59:37 +08:00
|
|
|
if (!PieceI) {
|
|
|
|
++I;
|
2013-05-04 09:13:01 +08:00
|
|
|
continue;
|
2013-05-07 05:59:37 +08:00
|
|
|
}
|
2013-05-04 09:13:01 +08:00
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s1Start = PieceI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s1End = PieceI->getEndLocation().getStmtOrNull();
|
2013-05-04 09:13:08 +08:00
|
|
|
const Stmt *level1 = getStmtParent(s1Start, PM);
|
|
|
|
const Stmt *level2 = getStmtParent(s1End, PM);
|
|
|
|
|
2013-05-04 09:13:22 +08:00
|
|
|
PathPieces::iterator NextI = I; ++NextI;
|
|
|
|
if (NextI == E)
|
|
|
|
break;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *PieceNextI = dyn_cast<PathDiagnosticControlFlowPiece>(NextI->get());
|
2013-05-04 09:13:01 +08:00
|
|
|
|
2013-05-07 05:59:37 +08:00
|
|
|
if (!PieceNextI) {
|
|
|
|
++I;
|
2013-05-04 09:13:01 +08:00
|
|
|
continue;
|
2013-05-07 05:59:37 +08:00
|
|
|
}
|
2013-05-04 09:13:01 +08:00
|
|
|
|
2018-06-13 04:51:01 +08:00
|
|
|
const Stmt *s2Start = PieceNextI->getStartLocation().getStmtOrNull();
|
|
|
|
const Stmt *s2End = PieceNextI->getEndLocation().getStmtOrNull();
|
2013-05-04 09:13:08 +08:00
|
|
|
const Stmt *level3 = getStmtParent(s2Start, PM);
|
|
|
|
const Stmt *level4 = getStmtParent(s2End, PM);
|
2013-05-04 09:13:01 +08:00
|
|
|
|
|
|
|
// Rule I.
|
|
|
|
//
|
|
|
|
// If we have two consecutive control edges whose end/begin locations
|
2013-05-07 05:59:37 +08:00
|
|
|
// are at the same level (e.g. statements or top-level expressions within
|
|
|
|
// a compound statement, or siblings share a single ancestor expression),
|
|
|
|
// then merge them if they have no interesting intermediate event.
|
2013-05-04 09:13:01 +08:00
|
|
|
//
|
|
|
|
// For example:
|
|
|
|
//
|
|
|
|
// (1.1 -> 1.2) -> (1.2 -> 1.3) becomes (1.1 -> 1.3) because the common
|
2013-05-07 05:59:37 +08:00
|
|
|
// parent is '1'. Here 'x.y.z' represents the hierarchy of statements.
|
2013-05-04 09:13:01 +08:00
|
|
|
//
|
|
|
|
// NOTE: this will be limited later in cases where we add barriers
|
|
|
|
// to prevent this optimization.
|
2013-05-07 05:59:37 +08:00
|
|
|
if (level1 && level1 == level2 && level1 == level3 && level1 == level4) {
|
2013-05-04 09:13:01 +08:00
|
|
|
PieceI->setEndLocation(PieceNextI->getEndLocation());
|
|
|
|
path.erase(NextI);
|
|
|
|
hasChanges = true;
|
|
|
|
continue;
|
|
|
|
}
|
2013-05-04 09:13:05 +08:00
|
|
|
|
|
|
|
// Rule II.
|
|
|
|
//
|
2013-05-07 15:30:07 +08:00
|
|
|
// Eliminate edges between subexpressions and parent expressions
|
|
|
|
// when the subexpression is consumed.
|
2013-05-04 09:13:08 +08:00
|
|
|
//
|
|
|
|
// NOTE: this will be limited later in cases where we add barriers
|
|
|
|
// to prevent this optimization.
|
2013-05-08 05:11:49 +08:00
|
|
|
if (s1End && s1End == s2Start && level2) {
|
2013-05-22 05:38:05 +08:00
|
|
|
bool removeEdge = false;
|
|
|
|
// Remove edges into the increment or initialization of a
|
|
|
|
// loop that have no interleaving event. This means that
|
|
|
|
// they aren't interesting.
|
|
|
|
if (isIncrementOrInitInForLoop(s1End, level2))
|
|
|
|
removeEdge = true;
|
|
|
|
// Next only consider edges that are not anchored on
|
|
|
|
// the condition of a terminator. This are intermediate edges
|
|
|
|
// that we might want to trim.
|
|
|
|
else if (!isConditionForTerminator(level2, s1End)) {
|
|
|
|
// Trim edges on expressions that are consumed by
|
|
|
|
// the parent expression.
|
|
|
|
if (isa<Expr>(s1End) && PM.isConsumedExpr(cast<Expr>(s1End))) {
|
2015-09-08 11:50:52 +08:00
|
|
|
removeEdge = true;
|
2013-05-22 05:38:05 +08:00
|
|
|
}
|
|
|
|
// Trim edges where a lexical containment doesn't exist.
|
|
|
|
// For example:
|
|
|
|
//
|
|
|
|
// X -> Y -> Z
|
|
|
|
//
|
|
|
|
// If 'Z' lexically contains Y (it is an ancestor) and
|
|
|
|
// 'X' does not lexically contain Y (it is a descendant OR
|
|
|
|
// it has no lexical relationship at all) then trim.
|
|
|
|
//
|
|
|
|
// This can eliminate edges where we dive into a subexpression
|
|
|
|
// and then pop back out, etc.
|
|
|
|
else if (s1Start && s2End &&
|
|
|
|
lexicalContains(PM, s2Start, s2End) &&
|
|
|
|
!lexicalContains(PM, s1End, s1Start)) {
|
|
|
|
removeEdge = true;
|
|
|
|
}
|
2013-06-04 07:00:09 +08:00
|
|
|
// Trim edges from a subexpression back to the top level if the
|
|
|
|
// subexpression is on a different line.
|
|
|
|
//
|
|
|
|
// A.1 -> A -> B
|
|
|
|
// becomes
|
|
|
|
// A.1 -> B
|
|
|
|
//
|
|
|
|
// These edges just look ugly and don't usually add anything.
|
|
|
|
else if (s1Start && s2End &&
|
|
|
|
lexicalContains(PM, s1Start, s1End)) {
|
|
|
|
SourceRange EdgeRange(PieceI->getEndLocation().asLocation(),
|
|
|
|
PieceI->getStartLocation().asLocation());
|
2022-06-21 01:51:34 +08:00
|
|
|
if (!getLengthOnSingleLine(SM, EdgeRange))
|
2013-06-04 07:00:09 +08:00
|
|
|
removeEdge = true;
|
|
|
|
}
|
2013-05-22 05:38:05 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (removeEdge) {
|
2013-05-08 05:11:49 +08:00
|
|
|
PieceI->setEndLocation(PieceNextI->getEndLocation());
|
|
|
|
path.erase(NextI);
|
|
|
|
hasChanges = true;
|
|
|
|
continue;
|
|
|
|
}
|
2013-05-04 09:13:08 +08:00
|
|
|
}
|
2013-05-07 05:59:37 +08:00
|
|
|
|
2013-05-21 08:34:40 +08:00
|
|
|
// Optimize edges for ObjC fast-enumeration loops.
|
|
|
|
//
|
|
|
|
// (X -> collection) -> (collection -> element)
|
|
|
|
//
|
|
|
|
// becomes:
|
|
|
|
//
|
|
|
|
// (X -> element)
|
|
|
|
if (s1End == s2Start) {
|
2018-03-07 08:17:48 +08:00
|
|
|
const auto *FS = dyn_cast_or_null<ObjCForCollectionStmt>(level3);
|
2013-05-21 08:34:40 +08:00
|
|
|
if (FS && FS->getCollection()->IgnoreParens() == s2Start &&
|
|
|
|
s2End == FS->getElement()) {
|
|
|
|
PieceI->setEndLocation(PieceNextI->getEndLocation());
|
|
|
|
path.erase(NextI);
|
|
|
|
hasChanges = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-05-07 05:59:37 +08:00
|
|
|
// No changes at this index? Move to the next one.
|
|
|
|
++I;
|
2013-05-04 09:13:01 +08:00
|
|
|
}
|
|
|
|
|
2013-05-23 02:52:35 +08:00
|
|
|
if (!hasChanges) {
|
2013-06-04 06:59:48 +08:00
|
|
|
// Adjust edges into subexpressions to make them more uniform
|
|
|
|
// and aesthetically pleasing.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addContextEdges(path, LC);
|
2013-06-04 07:00:00 +08:00
|
|
|
// Remove "cyclical" edges that include one or more context edges.
|
2018-09-29 02:49:41 +08:00
|
|
|
removeContextCycles(path, SM);
|
2013-06-04 06:59:48 +08:00
|
|
|
// Hoist edges originating from branch conditions to branches
|
|
|
|
// for simple branches.
|
|
|
|
simplifySimpleBranches(path);
|
2013-05-23 02:52:35 +08:00
|
|
|
// Remove any puny edges left over after primary optimization pass.
|
|
|
|
removePunyEdges(path, SM, PM);
|
2013-05-23 03:10:41 +08:00
|
|
|
// Remove identical events.
|
|
|
|
removeIdenticalEvents(path);
|
2013-05-23 02:52:35 +08:00
|
|
|
}
|
|
|
|
|
2013-05-04 09:13:01 +08:00
|
|
|
return hasChanges;
|
|
|
|
}
|
|
|
|
|
2013-06-04 06:59:45 +08:00
|
|
|
/// Drop the very first edge in a path, which should be a function entry edge.
|
2013-06-06 08:12:41 +08:00
|
|
|
///
|
|
|
|
/// If the first edge is not a function entry edge (say, because the first
|
|
|
|
/// statement had an invalid source location), this function does nothing.
|
|
|
|
// FIXME: We should just generate invalid edges anyway and have the optimizer
|
|
|
|
// deal with them.
|
2019-08-14 04:42:48 +08:00
|
|
|
static void dropFunctionEntryEdge(const PathDiagnosticConstruct &C,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathPieces &Path) {
|
2017-01-06 01:26:53 +08:00
|
|
|
const auto *FirstEdge =
|
|
|
|
dyn_cast<PathDiagnosticControlFlowPiece>(Path.front().get());
|
2013-06-06 08:12:41 +08:00
|
|
|
if (!FirstEdge)
|
|
|
|
return;
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const Decl *D = C.getLocationContextFor(&Path)->getDecl();
|
|
|
|
PathDiagnosticLocation EntryLoc =
|
|
|
|
PathDiagnosticLocation::createBegin(D, C.getSourceManager());
|
2013-06-06 08:12:41 +08:00
|
|
|
if (FirstEdge->getStartLocation() != EntryLoc)
|
|
|
|
return;
|
2013-06-04 06:59:45 +08:00
|
|
|
|
|
|
|
Path.pop_front();
|
|
|
|
}
|
|
|
|
|
2018-09-07 08:42:53 +08:00
|
|
|
/// Populate executes lines with lines containing at least one diagnostics.
|
2019-08-14 00:45:48 +08:00
|
|
|
static void updateExecutedLinesWithDiagnosticPieces(PathDiagnostic &PD) {
|
2018-09-07 08:42:53 +08:00
|
|
|
|
|
|
|
PathPieces path = PD.path.flatten(/*ShouldFlattenMacros=*/true);
|
|
|
|
FilesToLineNumsMap &ExecutedLines = PD.getExecutedLines();
|
|
|
|
|
|
|
|
for (const auto &P : path) {
|
|
|
|
FullSourceLoc Loc = P->getLocation().asLocation().getExpansionLoc();
|
|
|
|
FileID FID = Loc.getFileID();
|
|
|
|
unsigned LineNo = Loc.getLineNumber();
|
2018-09-07 08:43:17 +08:00
|
|
|
assert(FID.isValid());
|
2018-09-07 08:43:37 +08:00
|
|
|
ExecutedLines[FID].insert(LineNo);
|
2018-09-07 08:42:53 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticConstruct::PathDiagnosticConstruct(
|
|
|
|
const PathDiagnosticConsumer *PDC, const ExplodedNode *ErrorNode,
|
2019-09-10 04:34:40 +08:00
|
|
|
const PathSensitiveBugReport *R)
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
: Consumer(PDC), CurrentNode(ErrorNode),
|
|
|
|
SM(CurrentNode->getCodeDecl().getASTContext().getSourceManager()),
|
|
|
|
PD(generateEmptyDiagnosticForReport(R, getSourceManager())) {
|
|
|
|
LCM[&PD->getActivePath()] = ErrorNode->getLocationContext();
|
|
|
|
}
|
|
|
|
|
|
|
|
PathDiagnosticBuilder::PathDiagnosticBuilder(
|
|
|
|
BugReporterContext BRC, std::unique_ptr<ExplodedGraph> BugPath,
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport *r, const ExplodedNode *ErrorNode,
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
std::unique_ptr<VisitorsDiagnosticsTy> VisitorsDiagnostics)
|
|
|
|
: BugReporterContext(BRC), BugPath(std::move(BugPath)), R(r),
|
|
|
|
ErrorNode(ErrorNode),
|
|
|
|
VisitorsDiagnostics(std::move(VisitorsDiagnostics)) {}
|
|
|
|
|
|
|
|
std::unique_ptr<PathDiagnostic>
|
|
|
|
PathDiagnosticBuilder::generate(const PathDiagnosticConsumer *PDC) const {
|
2019-08-14 04:42:48 +08:00
|
|
|
PathDiagnosticConstruct Construct(PDC, ErrorNode, R);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
|
|
|
const SourceManager &SM = getSourceManager();
|
|
|
|
const AnalyzerOptions &Opts = getAnalyzerOptions();
|
2019-08-16 09:53:14 +08:00
|
|
|
|
|
|
|
if (!PDC->shouldGenerateDiagnostics())
|
|
|
|
return generateEmptyDiagnosticForReport(R, getSourceManager());
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
|
|
|
|
// Construct the final (warning) event for the bug report.
|
|
|
|
auto EndNotes = VisitorsDiagnostics->find(ErrorNode);
|
|
|
|
PathDiagnosticPieceRef LastPiece;
|
|
|
|
if (EndNotes != VisitorsDiagnostics->end()) {
|
|
|
|
assert(!EndNotes->second.empty());
|
|
|
|
LastPiece = EndNotes->second[0];
|
|
|
|
} else {
|
|
|
|
LastPiece = BugReporterVisitor::getDefaultEndPath(*this, ErrorNode,
|
|
|
|
*getBugReport());
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Construct.PD->setEndOfPath(LastPiece);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
PathDiagnosticLocation PrevLoc = Construct.PD->getLocation();
|
|
|
|
// From the error node to the root, ascend the bug path and construct the bug
|
|
|
|
// report.
|
|
|
|
while (Construct.ascendToPrevNode()) {
|
|
|
|
generatePathDiagnosticsForNode(Construct, PrevLoc);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
auto VisitorNotes = VisitorsDiagnostics->find(Construct.getCurrentNode());
|
|
|
|
if (VisitorNotes == VisitorsDiagnostics->end())
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// This is a workaround due to inability to put shared PathDiagnosticPiece
|
|
|
|
// into a FoldingSet.
|
|
|
|
std::set<llvm::FoldingSetNodeID> DeduplicationSet;
|
|
|
|
|
|
|
|
// Add pieces from custom visitors.
|
2019-08-14 00:45:48 +08:00
|
|
|
for (const PathDiagnosticPieceRef &Note : VisitorNotes->second) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
llvm::FoldingSetNodeID ID;
|
|
|
|
Note->Profile(ID);
|
2019-08-14 02:48:08 +08:00
|
|
|
if (!DeduplicationSet.insert(ID).second)
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
continue;
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (PDC->shouldAddPathEdges())
|
|
|
|
addEdgeToPath(Construct.getActivePath(), PrevLoc, Note->getLocation());
|
2019-09-12 04:54:17 +08:00
|
|
|
updateStackPiecesWithMessage(Note, Construct.CallStack);
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Construct.getActivePath().push_front(Note);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (PDC->shouldAddPathEdges()) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Add an edge to the start of the function.
|
|
|
|
// We'll prune it out later, but it helps make diagnostics more uniform.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
const StackFrameContext *CalleeLC =
|
|
|
|
Construct.getLocationContextForActivePath()->getStackFrame();
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
const Decl *D = CalleeLC->getDecl();
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
addEdgeToPath(Construct.getActivePath(), PrevLoc,
|
2018-09-29 02:49:41 +08:00
|
|
|
PathDiagnosticLocation::createBegin(D, SM));
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// Finally, prune the diagnostic path of uninteresting stuff.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (!Construct.PD->path.empty()) {
|
[analyzer] Evaluate all non-checker config options before analysis
In earlier patches regarding AnalyzerOptions, a lot of effort went into
gathering all config options, and changing the interface so that potential
misuse can be eliminited.
Up until this point, AnalyzerOptions only evaluated an option when it was
querried. For example, if we had a "-no-false-positives" flag, AnalyzerOptions
would store an Optional field for it that would be None up until somewhere in
the code until the flag's getter function is called.
However, now that we're confident that we've gathered all configs, we can
evaluate off of them before analysis, so we can emit a error on invalid input
even if that prticular flag will not matter in that particular run of the
analyzer. Another very big benefit of this is that debug.ConfigDumper will now
show the value of all configs every single time.
Also, almost all options related class have a similar interface, so uniformity
is also a benefit.
The implementation for errors on invalid input will be commited shorty.
Differential Revision: https://reviews.llvm.org/D53692
llvm-svn: 348031
2018-12-01 04:44:00 +08:00
|
|
|
if (R->shouldPrunePath() && Opts.ShouldPrunePaths) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
bool stillHasNotes =
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
removeUnneededCalls(Construct, Construct.getMutablePieces(), R);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
assert(stillHasNotes);
|
|
|
|
(void)stillHasNotes;
|
|
|
|
}
|
|
|
|
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
// Remove pop-up notes if needed.
|
|
|
|
if (!Opts.ShouldAddPopUpNotes)
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
removePopUpNotes(Construct.getMutablePieces());
|
[analyzer] [NFC] PathDiagnostic: Create PathDiagnosticPopUpPiece
Summary:
This new piece is similar to our macro expansion printing in HTML reports:
On mouse-hover event it pops up on variables. Similar to note pieces it
supports `plist` diagnostics as well.
It is optional, on by default: `add-pop-up-notes=true`.
Extra: In HTML reports `background-color: LemonChiffon` was too light,
changed to `PaleGoldenRod`.
Reviewers: NoQ, alexfh
Reviewed By: NoQ
Subscribers: cfe-commits, gerazo, gsd, george.karpenkov, alexfh, xazax.hun,
baloghadamsoftware, szepet, a.sidorin, mikhail.ramalho,
Szelethus, donat.nagy, dkrupp
Tags: #clang
Differential Revision: https://reviews.llvm.org/D60670
llvm-svn: 362014
2019-05-30 03:21:59 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Redirect all call pieces to have valid locations.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
adjustCallLocations(Construct.getMutablePieces());
|
|
|
|
removePiecesWithInvalidLocations(Construct.getMutablePieces());
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (PDC->shouldAddPathEdges()) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
|
|
|
// Reduce the number of edges from a very conservative set
|
|
|
|
// to an aesthetically pleasing subset that conveys the
|
|
|
|
// necessary information.
|
|
|
|
OptimizedCallsSet OCS;
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
while (optimizeEdges(Construct, Construct.getMutablePieces(), OCS)) {
|
|
|
|
}
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
|
|
|
// Drop the very first function-entry edge. It's not really necessary
|
|
|
|
// for top-level functions.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
dropFunctionEntryEdge(Construct, Construct.getMutablePieces());
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// Remove messages that are basically the same, and edges that may not
|
|
|
|
// make sense.
|
|
|
|
// We have to do this after edge optimization in the Extensive mode.
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
removeRedundantMsgs(Construct.getMutablePieces());
|
|
|
|
removeEdgesToDefaultInitializers(Construct.getMutablePieces());
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
2018-10-31 22:54:27 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
if (Opts.ShouldDisplayMacroExpansions)
|
|
|
|
CompactMacroExpandedPieces(Construct.getMutablePieces(), SM);
|
2018-10-31 22:54:27 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
return std::move(Construct.PD);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Methods for BugType and subclasses.
|
|
|
|
//===----------------------------------------------------------------------===//
|
2018-03-07 08:17:48 +08:00
|
|
|
|
|
|
|
void BugType::anchor() {}
|
2011-02-23 08:16:01 +08:00
|
|
|
|
2011-12-20 10:48:34 +08:00
|
|
|
void BuiltinBug::anchor() {}
|
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Methods for BugReport and subclasses.
|
|
|
|
//===----------------------------------------------------------------------===//
|
2011-08-18 07:00:25 +08:00
|
|
|
|
2020-05-27 18:27:32 +08:00
|
|
|
LLVM_ATTRIBUTE_USED static bool
|
|
|
|
isDependency(const CheckerRegistryData &Registry, StringRef CheckerName) {
|
2020-05-27 18:27:32 +08:00
|
|
|
for (const std::pair<StringRef, StringRef> &Pair : Registry.Dependencies) {
|
|
|
|
if (Pair.second == CheckerName)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2020-05-27 18:27:32 +08:00
|
|
|
LLVM_ATTRIBUTE_USED static bool isHidden(const CheckerRegistryData &Registry,
|
|
|
|
StringRef CheckerName) {
|
|
|
|
for (const CheckerInfo &Checker : Registry.Checkers) {
|
|
|
|
if (Checker.FullName == CheckerName)
|
|
|
|
return Checker.IsHidden;
|
|
|
|
}
|
|
|
|
llvm_unreachable(
|
|
|
|
"Checker name not found in CheckerRegistry -- did you retrieve it "
|
|
|
|
"correctly from CheckerManager::getCurrentCheckerName?");
|
|
|
|
}
|
|
|
|
|
2020-05-27 18:27:32 +08:00
|
|
|
PathSensitiveBugReport::PathSensitiveBugReport(
|
|
|
|
const BugType &bt, StringRef shortDesc, StringRef desc,
|
|
|
|
const ExplodedNode *errorNode, PathDiagnosticLocation LocationToUnique,
|
|
|
|
const Decl *DeclToUnique)
|
|
|
|
: BugReport(Kind::PathSensitive, bt, shortDesc, desc), ErrorNode(errorNode),
|
|
|
|
ErrorNodeRange(getStmt() ? getStmt()->getSourceRange() : SourceRange()),
|
|
|
|
UniqueingLocation(LocationToUnique), UniqueingDecl(DeclToUnique) {
|
|
|
|
assert(!isDependency(ErrorNode->getState()
|
|
|
|
->getAnalysisManager()
|
|
|
|
.getCheckerManager()
|
|
|
|
->getCheckerRegistryData(),
|
|
|
|
bt.getCheckerName()) &&
|
|
|
|
"Some checkers depend on this one! We don't allow dependency "
|
|
|
|
"checkers to emit warnings, because checkers should depend on "
|
|
|
|
"*modeling*, not *diagnostics*.");
|
2020-05-27 18:27:32 +08:00
|
|
|
|
|
|
|
assert(
|
2020-07-08 16:14:53 +08:00
|
|
|
(bt.getCheckerName().startswith("debug") ||
|
|
|
|
!isHidden(ErrorNode->getState()
|
|
|
|
->getAnalysisManager()
|
|
|
|
.getCheckerManager()
|
|
|
|
->getCheckerRegistryData(),
|
|
|
|
bt.getCheckerName())) &&
|
2020-05-27 18:27:32 +08:00
|
|
|
"Hidden checkers musn't emit diagnostics as they are by definition "
|
|
|
|
"non-user facing!");
|
2020-05-27 18:27:32 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::addVisitor(
|
|
|
|
std::unique_ptr<BugReporterVisitor> visitor) {
|
2011-08-19 06:37:56 +08:00
|
|
|
if (!visitor)
|
|
|
|
return;
|
|
|
|
|
|
|
|
llvm::FoldingSetNodeID ID;
|
|
|
|
visitor->Profile(ID);
|
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
void *InsertPos = nullptr;
|
|
|
|
if (CallbacksSet.FindNodeOrInsertPos(ID, InsertPos)) {
|
2011-08-19 06:37:56 +08:00
|
|
|
return;
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
2011-08-19 06:37:56 +08:00
|
|
|
|
2014-09-05 07:54:33 +08:00
|
|
|
Callbacks.push_back(std::move(visitor));
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::clearVisitors() {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
Callbacks.clear();
|
2011-08-19 06:37:56 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
const Decl *PathSensitiveBugReport::getDeclWithIssue() const {
|
2012-04-05 02:11:35 +08:00
|
|
|
const ExplodedNode *N = getErrorNode();
|
|
|
|
if (!N)
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
|
|
|
|
2012-04-05 02:11:35 +08:00
|
|
|
const LocationContext *LC = N->getLocationContext();
|
2018-06-27 09:51:55 +08:00
|
|
|
return LC->getStackFrame()->getDecl();
|
2012-04-05 02:11:35 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void BasicBugReport::Profile(llvm::FoldingSetNodeID& hash) const {
|
|
|
|
hash.AddInteger(static_cast<int>(getKind()));
|
|
|
|
hash.AddPointer(&BT);
|
|
|
|
hash.AddString(Description);
|
|
|
|
assert(Location.isValid());
|
|
|
|
Location.Profile(hash);
|
|
|
|
|
|
|
|
for (SourceRange range : Ranges) {
|
|
|
|
if (!range.isValid())
|
|
|
|
continue;
|
2020-10-27 18:43:39 +08:00
|
|
|
hash.Add(range.getBegin());
|
|
|
|
hash.Add(range.getEnd());
|
2019-09-10 04:34:40 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void PathSensitiveBugReport::Profile(llvm::FoldingSetNodeID &hash) const {
|
|
|
|
hash.AddInteger(static_cast<int>(getKind()));
|
2011-08-18 07:00:25 +08:00
|
|
|
hash.AddPointer(&BT);
|
|
|
|
hash.AddString(Description);
|
2013-01-08 08:25:29 +08:00
|
|
|
PathDiagnosticLocation UL = getUniqueingLocation();
|
|
|
|
if (UL.isValid()) {
|
|
|
|
UL.Profile(hash);
|
2011-09-21 05:38:35 +08:00
|
|
|
} else {
|
2019-09-12 04:54:21 +08:00
|
|
|
// TODO: The statement may be null if the report was emitted before any
|
|
|
|
// statements were executed. In particular, some checkers by design
|
|
|
|
// occasionally emit their reports in empty functions (that have no
|
|
|
|
// statements in their body). Do we profile correctly in this case?
|
|
|
|
hash.AddPointer(ErrorNode->getCurrentOrPreviousStmtForDiagnostics());
|
2011-09-21 05:38:35 +08:00
|
|
|
}
|
2011-08-18 07:00:25 +08:00
|
|
|
|
2015-10-04 12:53:55 +08:00
|
|
|
for (SourceRange range : Ranges) {
|
2011-08-18 07:00:25 +08:00
|
|
|
if (!range.isValid())
|
|
|
|
continue;
|
2020-10-27 18:43:39 +08:00
|
|
|
hash.Add(range.getBegin());
|
|
|
|
hash.Add(range.getEnd());
|
2011-08-18 07:00:25 +08:00
|
|
|
}
|
|
|
|
}
|
2009-04-01 07:00:32 +08:00
|
|
|
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
template <class T>
|
|
|
|
static void insertToInterestingnessMap(
|
|
|
|
llvm::DenseMap<T, bugreporter::TrackingKind> &InterestingnessMap, T Val,
|
|
|
|
bugreporter::TrackingKind TKind) {
|
|
|
|
auto Result = InterestingnessMap.insert({Val, TKind});
|
|
|
|
|
|
|
|
if (Result.second)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Even if this symbol/region was already marked as interesting as a
|
|
|
|
// condition, if we later mark it as interesting again but with
|
|
|
|
// thorough tracking, overwrite it. Entities marked with thorough
|
|
|
|
// interestiness are the most important (or most interesting, if you will),
|
|
|
|
// and we wouldn't like to downplay their importance.
|
|
|
|
|
|
|
|
switch (TKind) {
|
|
|
|
case bugreporter::TrackingKind::Thorough:
|
|
|
|
Result.first->getSecond() = bugreporter::TrackingKind::Thorough;
|
|
|
|
return;
|
|
|
|
case bugreporter::TrackingKind::Condition:
|
|
|
|
return;
|
2020-05-27 18:27:32 +08:00
|
|
|
}
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
|
2020-05-27 18:27:32 +08:00
|
|
|
llvm_unreachable(
|
|
|
|
"BugReport::markInteresting currently can only handle 2 different "
|
|
|
|
"tracking kinds! Please define what tracking kind should this entitiy"
|
|
|
|
"have, if it was already marked as interesting with a different kind!");
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::markInteresting(SymbolRef sym,
|
|
|
|
bugreporter::TrackingKind TKind) {
|
2012-03-09 09:13:14 +08:00
|
|
|
if (!sym)
|
|
|
|
return;
|
2012-03-24 11:03:29 +08:00
|
|
|
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
insertToInterestingnessMap(InterestingSymbols, sym, TKind);
|
2012-03-16 06:45:29 +08:00
|
|
|
|
2021-07-15 14:34:59 +08:00
|
|
|
// FIXME: No tests exist for this code and it is questionable:
|
|
|
|
// How to handle multiple metadata for the same region?
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *meta = dyn_cast<SymbolMetadata>(sym))
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
markInteresting(meta->getRegion(), TKind);
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
|
|
|
|
2021-07-15 14:34:59 +08:00
|
|
|
void PathSensitiveBugReport::markNotInteresting(SymbolRef sym) {
|
|
|
|
if (!sym)
|
|
|
|
return;
|
|
|
|
InterestingSymbols.erase(sym);
|
|
|
|
|
|
|
|
// The metadata part of markInteresting is not reversed here.
|
|
|
|
// Just making the same region not interesting is incorrect
|
|
|
|
// in specific cases.
|
2021-07-18 17:10:31 +08:00
|
|
|
if (const auto *meta = dyn_cast<SymbolMetadata>(sym))
|
|
|
|
markNotInteresting(meta->getRegion());
|
2021-07-15 14:34:59 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::markInteresting(const MemRegion *R,
|
|
|
|
bugreporter::TrackingKind TKind) {
|
2012-03-09 09:13:14 +08:00
|
|
|
if (!R)
|
|
|
|
return;
|
2012-03-24 11:03:29 +08:00
|
|
|
|
2012-03-09 09:13:14 +08:00
|
|
|
R = R->getBaseRegion();
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
insertToInterestingnessMap(InterestingRegions, R, TKind);
|
2012-03-16 06:45:29 +08:00
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *SR = dyn_cast<SymbolicRegion>(R))
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
markInteresting(SR->getSymbol(), TKind);
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
|
|
|
|
2021-07-15 14:34:59 +08:00
|
|
|
void PathSensitiveBugReport::markNotInteresting(const MemRegion *R) {
|
|
|
|
if (!R)
|
|
|
|
return;
|
|
|
|
|
|
|
|
R = R->getBaseRegion();
|
|
|
|
InterestingRegions.erase(R);
|
|
|
|
|
|
|
|
if (const auto *SR = dyn_cast<SymbolicRegion>(R))
|
|
|
|
markNotInteresting(SR->getSymbol());
|
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::markInteresting(SVal V,
|
|
|
|
bugreporter::TrackingKind TKind) {
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
markInteresting(V.getAsRegion(), TKind);
|
|
|
|
markInteresting(V.getAsSymbol(), TKind);
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReport::markInteresting(const LocationContext *LC) {
|
2012-08-30 05:22:37 +08:00
|
|
|
if (!LC)
|
|
|
|
return;
|
|
|
|
InterestingLocationContexts.insert(LC);
|
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
Optional<bugreporter::TrackingKind>
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport::getInterestingnessKind(SVal V) const {
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
auto RKind = getInterestingnessKind(V.getAsRegion());
|
|
|
|
auto SKind = getInterestingnessKind(V.getAsSymbol());
|
|
|
|
if (!RKind)
|
|
|
|
return SKind;
|
|
|
|
if (!SKind)
|
|
|
|
return RKind;
|
|
|
|
|
|
|
|
// If either is marked with throrough tracking, return that, we wouldn't like
|
|
|
|
// to downplay a note's importance by 'only' mentioning it as a condition.
|
|
|
|
switch(*RKind) {
|
|
|
|
case bugreporter::TrackingKind::Thorough:
|
|
|
|
return RKind;
|
|
|
|
case bugreporter::TrackingKind::Condition:
|
|
|
|
return SKind;
|
|
|
|
}
|
|
|
|
|
|
|
|
llvm_unreachable(
|
|
|
|
"BugReport::getInterestingnessKind currently can only handle 2 different "
|
|
|
|
"tracking kinds! Please define what tracking kind should we return here "
|
|
|
|
"when the kind of getAsRegion() and getAsSymbol() is different!");
|
|
|
|
return None;
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
Optional<bugreporter::TrackingKind>
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport::getInterestingnessKind(SymbolRef sym) const {
|
2012-03-09 09:13:14 +08:00
|
|
|
if (!sym)
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
return None;
|
2012-03-16 06:45:29 +08:00
|
|
|
// We don't currently consider metadata symbols to be interesting
|
|
|
|
// even if we know their region is interesting. Is that correct behavior?
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
auto It = InterestingSymbols.find(sym);
|
|
|
|
if (It == InterestingSymbols.end())
|
|
|
|
return None;
|
|
|
|
return It->getSecond();
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
Optional<bugreporter::TrackingKind>
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport::getInterestingnessKind(const MemRegion *R) const {
|
2012-03-09 09:13:14 +08:00
|
|
|
if (!R)
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
return None;
|
|
|
|
|
2012-03-09 09:13:14 +08:00
|
|
|
R = R->getBaseRegion();
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
auto It = InterestingRegions.find(R);
|
|
|
|
if (It != InterestingRegions.end())
|
|
|
|
return It->getSecond();
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *SR = dyn_cast<SymbolicRegion>(R))
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
return getInterestingnessKind(SR->getSymbol());
|
|
|
|
return None;
|
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
bool PathSensitiveBugReport::isInteresting(SVal V) const {
|
2022-06-26 02:56:50 +08:00
|
|
|
return getInterestingnessKind(V).hasValue();
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
bool PathSensitiveBugReport::isInteresting(SymbolRef sym) const {
|
2022-06-26 02:56:50 +08:00
|
|
|
return getInterestingnessKind(sym).hasValue();
|
[analyzer][NFC] Add different interestingness kinds
We defined (on the mailing list and here on phabricator) 2 different cases where
retrieving information about a control dependency condition is very important:
* When the condition's last write happened in a different stack frame
* When the collapse point of the condition (when we can constrain it to be
true/false) didn't happen in the actual condition.
It seems like we solved this problem with the help of expression value tracking,
and have started working on better diagnostics notes about this process.
Expression value tracking is nothing more than registering a variety of visitors
to construct reports about it. Each of the registered visitors (ReturnVisitor,
FindLastStoreVisitor, NoStoreFuncVisitor, etc) have something to go by: a
MemRegion, an SVal, an ExplodedNode, etc. For this reason, better explaining a
last write is super simple, we can always just pass on some more information to
the visitor in question (as seen in D65575).
ConditionBRVisitor is a different beast, as it was built for a different
purpose. It is responsible for constructing events at, well, conditions, and is
registered only once, and isn't a part of the "expression value tracking
family". Unfortunately, it is also the visitor to tinker with for constructing
better diagnostics about the collapse point problem.
This creates a need for alternative way to communicate with ConditionBRVisitor
that a specific condition is being tracked for for the reason of being a control
dependency. Since at almost all PathDiagnosticEventPiece construction the
visitor checks interestingness, it makes sense to pair interestingness with a
reason as to why we marked an entity as such.
Differential Revision: https://reviews.llvm.org/D65723
llvm-svn: 369583
2019-08-22 05:33:25 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
bool PathSensitiveBugReport::isInteresting(const MemRegion *R) const {
|
2022-06-26 02:56:50 +08:00
|
|
|
return getInterestingnessKind(R).hasValue();
|
2012-03-09 09:13:14 +08:00
|
|
|
}
|
Allow multiple PathDiagnosticConsumers to be used with a BugReporter at the same time.
This fixes several issues:
- removes egregious hack where PlistDiagnosticConsumer would forward to HTMLDiagnosticConsumer,
but diagnostics wouldn't be generated consistently in the same way if PlistDiagnosticConsumer
was used by itself.
- emitting diagnostics to the terminal (using clang's diagnostic machinery) is no longer a special
case, just another PathDiagnosticConsumer. This also magically resolved some duplicate warnings,
as we now use PathDiagnosticConsumer's diagnostic pruning, which has scope for the entire translation
unit, not just the scope of a BugReporter (which is limited to a particular ExprEngine).
As an interesting side-effect, diagnostics emitted to the terminal also have their trailing "." stripped,
just like with diagnostics emitted to plists and HTML. This required some tests to be updated, but now
the tests have higher fidelity with what users will see.
There are some inefficiencies in this patch. We currently generate the report graph (from the ExplodedGraph)
once per PathDiagnosticConsumer, which is a bit wasteful, but that could be pulled up higher in the
logic stack. There is some intended duplication, however, as we now generate different PathDiagnostics (for the same issue)
for different PathDiagnosticConsumers. This is necessary to produce the diagnostics that a particular
consumer expects.
llvm-svn: 162028
2012-08-17 01:45:23 +08:00
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
bool PathSensitiveBugReport::isInteresting(const LocationContext *LC) const {
|
2012-08-30 05:22:37 +08:00
|
|
|
if (!LC)
|
|
|
|
return false;
|
|
|
|
return InterestingLocationContexts.count(LC);
|
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
const Stmt *PathSensitiveBugReport::getStmt() const {
|
2011-08-18 07:00:25 +08:00
|
|
|
if (!ErrorNode)
|
2014-05-27 10:45:47 +08:00
|
|
|
return nullptr;
|
2011-08-18 07:00:25 +08:00
|
|
|
|
2010-09-16 11:50:38 +08:00
|
|
|
ProgramPoint ProgP = ErrorNode->getLocation();
|
2014-05-27 10:45:47 +08:00
|
|
|
const Stmt *S = nullptr;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-02-22 06:23:56 +08:00
|
|
|
if (Optional<BlockEntrance> BE = ProgP.getAs<BlockEntrance>()) {
|
2009-08-20 09:23:34 +08:00
|
|
|
CFGBlock &Exit = ProgP.getLocationContext()->getCFG()->getExit();
|
2009-08-18 16:46:04 +08:00
|
|
|
if (BE->getBlock() == &Exit)
|
2019-09-12 04:54:21 +08:00
|
|
|
S = ErrorNode->getPreviousStmtForDiagnostics();
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
2009-07-23 06:35:28 +08:00
|
|
|
if (!S)
|
2019-09-12 04:54:21 +08:00
|
|
|
S = ErrorNode->getStmtForDiagnostics();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
|
|
|
return S;
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
ArrayRef<SourceRange>
|
|
|
|
PathSensitiveBugReport::getRanges() const {
|
2015-02-07 01:25:10 +08:00
|
|
|
// If no custom ranges, add the range of the statement corresponding to
|
|
|
|
// the error node.
|
2019-09-10 04:34:40 +08:00
|
|
|
if (Ranges.empty() && isa_and_nonnull<Expr>(getStmt()))
|
|
|
|
return ErrorNodeRange;
|
2011-08-18 07:00:25 +08:00
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
return Ranges;
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
PathDiagnosticLocation
|
|
|
|
PathSensitiveBugReport::getLocation() const {
|
2019-09-12 04:54:21 +08:00
|
|
|
assert(ErrorNode && "Cannot create a location with a null node.");
|
|
|
|
const Stmt *S = ErrorNode->getStmtForDiagnostics();
|
|
|
|
ProgramPoint P = ErrorNode->getLocation();
|
|
|
|
const LocationContext *LC = P.getLocationContext();
|
|
|
|
SourceManager &SM =
|
|
|
|
ErrorNode->getState()->getStateManager().getContext().getSourceManager();
|
|
|
|
|
|
|
|
if (!S) {
|
|
|
|
// If this is an implicit call, return the implicit call point location.
|
|
|
|
if (Optional<PreImplicitCall> PIE = P.getAs<PreImplicitCall>())
|
|
|
|
return PathDiagnosticLocation(PIE->getLocation(), SM);
|
|
|
|
if (auto FE = P.getAs<FunctionExitPoint>()) {
|
|
|
|
if (const ReturnStmt *RS = FE->getStmt())
|
|
|
|
return PathDiagnosticLocation::createBegin(RS, SM, LC);
|
|
|
|
}
|
|
|
|
S = ErrorNode->getNextStmtForDiagnostics();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (S) {
|
|
|
|
// For member expressions, return the location of the '.' or '->'.
|
|
|
|
if (const auto *ME = dyn_cast<MemberExpr>(S))
|
|
|
|
return PathDiagnosticLocation::createMemberLoc(ME, SM);
|
|
|
|
|
|
|
|
// For binary operators, return the location of the operator.
|
|
|
|
if (const auto *B = dyn_cast<BinaryOperator>(S))
|
|
|
|
return PathDiagnosticLocation::createOperatorLoc(B, SM);
|
|
|
|
|
|
|
|
if (P.getAs<PostStmtPurgeDeadSymbols>())
|
|
|
|
return PathDiagnosticLocation::createEnd(S, SM, LC);
|
|
|
|
|
|
|
|
if (S->getBeginLoc().isValid())
|
|
|
|
return PathDiagnosticLocation(S, SM, LC);
|
|
|
|
|
|
|
|
return PathDiagnosticLocation(
|
|
|
|
PathDiagnosticLocation::getValidSourceLocation(S, LC), SM);
|
|
|
|
}
|
|
|
|
|
|
|
|
return PathDiagnosticLocation::createDeclEnd(ErrorNode->getLocationContext(),
|
|
|
|
SM);
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Methods for BugReporter and subclasses.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2019-08-20 10:15:50 +08:00
|
|
|
const ExplodedGraph &PathSensitiveBugReporter::getGraph() const {
|
|
|
|
return Eng.getGraph();
|
|
|
|
}
|
2009-04-01 07:00:32 +08:00
|
|
|
|
2019-08-20 10:15:50 +08:00
|
|
|
ProgramStateManager &PathSensitiveBugReporter::getStateManager() const {
|
|
|
|
return Eng.getStateManager();
|
|
|
|
}
|
2019-08-14 02:48:08 +08:00
|
|
|
|
2020-04-29 10:14:58 +08:00
|
|
|
BugReporter::BugReporter(BugReporterData &d) : D(d) {}
|
2011-08-19 09:57:09 +08:00
|
|
|
BugReporter::~BugReporter() {
|
2019-08-21 05:41:20 +08:00
|
|
|
// Make sure reports are flushed.
|
|
|
|
assert(StrBugTypes.empty() &&
|
|
|
|
"Destroying BugReporter before diagnostics are emitted!");
|
2011-08-19 09:57:09 +08:00
|
|
|
|
|
|
|
// Free the bug reports we are tracking.
|
2018-03-07 08:17:48 +08:00
|
|
|
for (const auto I : EQClassesVector)
|
|
|
|
delete I;
|
2011-08-19 09:57:09 +08:00
|
|
|
}
|
2009-04-01 07:00:32 +08:00
|
|
|
|
|
|
|
void BugReporter::FlushReports() {
|
2012-08-03 07:41:05 +08:00
|
|
|
// We need to flush reports in deterministic order to ensure the order
|
|
|
|
// of the reports is consistent between runs.
|
2018-03-07 08:17:48 +08:00
|
|
|
for (const auto EQ : EQClassesVector)
|
|
|
|
FlushReport(*EQ);
|
2009-04-01 07:00:32 +08:00
|
|
|
|
2011-02-23 08:16:01 +08:00
|
|
|
// BugReporter owns and deletes only BugTypes created implicitly through
|
|
|
|
// EmitBasicReport.
|
|
|
|
// FIXME: There are leaks from checkers that assume that the BugTypes they
|
|
|
|
// create will be destroyed by the BugReporter.
|
2020-04-29 10:14:58 +08:00
|
|
|
StrBugTypes.clear();
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// PathDiagnostics generation.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2013-03-16 09:07:58 +08:00
|
|
|
namespace {
|
2018-03-07 08:17:48 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
/// A wrapper around an ExplodedGraph that contains a single path from the root
|
2019-08-21 16:48:24 +08:00
|
|
|
/// to the error node.
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
class BugPathInfo {
|
2013-03-16 09:07:58 +08:00
|
|
|
public:
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
std::unique_ptr<ExplodedGraph> BugPath;
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport *Report;
|
2013-03-23 05:15:28 +08:00
|
|
|
const ExplodedNode *ErrorNode;
|
2013-03-16 09:07:58 +08:00
|
|
|
};
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
/// A wrapper around an ExplodedGraph whose leafs are all error nodes. Can
|
|
|
|
/// conveniently retrieve bug paths from a single error node to the root.
|
|
|
|
class BugPathGetter {
|
|
|
|
std::unique_ptr<ExplodedGraph> TrimmedGraph;
|
|
|
|
|
2018-03-07 08:17:48 +08:00
|
|
|
using PriorityMapTy = llvm::DenseMap<const ExplodedNode *, unsigned>;
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
/// Assign each node with its distance from the root.
|
2013-03-20 08:35:37 +08:00
|
|
|
PriorityMapTy PriorityMap;
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
/// Since the getErrorNode() or BugReport refers to the original ExplodedGraph,
|
|
|
|
/// we need to pair it to the error node of the constructed trimmed graph.
|
2019-09-10 04:34:40 +08:00
|
|
|
using ReportNewNodePair =
|
|
|
|
std::pair<PathSensitiveBugReport *, const ExplodedNode *>;
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
SmallVector<ReportNewNodePair, 32> ReportNodes;
|
2013-03-23 05:15:28 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
BugPathInfo CurrentBugPath;
|
2013-03-23 05:15:28 +08:00
|
|
|
|
|
|
|
/// A helper class for sorting ExplodedNodes by priority.
|
|
|
|
template <bool Descending>
|
|
|
|
class PriorityCompare {
|
|
|
|
const PriorityMapTy &PriorityMap;
|
|
|
|
|
|
|
|
public:
|
|
|
|
PriorityCompare(const PriorityMapTy &M) : PriorityMap(M) {}
|
|
|
|
|
|
|
|
bool operator()(const ExplodedNode *LHS, const ExplodedNode *RHS) const {
|
|
|
|
PriorityMapTy::const_iterator LI = PriorityMap.find(LHS);
|
|
|
|
PriorityMapTy::const_iterator RI = PriorityMap.find(RHS);
|
|
|
|
PriorityMapTy::const_iterator E = PriorityMap.end();
|
|
|
|
|
|
|
|
if (LI == E)
|
|
|
|
return Descending;
|
|
|
|
if (RI == E)
|
|
|
|
return !Descending;
|
|
|
|
|
|
|
|
return Descending ? LI->second > RI->second
|
|
|
|
: LI->second < RI->second;
|
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
bool operator()(const ReportNewNodePair &LHS,
|
|
|
|
const ReportNewNodePair &RHS) const {
|
|
|
|
return (*this)(LHS.second, RHS.second);
|
2013-03-23 05:15:28 +08:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2013-03-16 09:07:58 +08:00
|
|
|
public:
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
BugPathGetter(const ExplodedGraph *OriginalGraph,
|
2019-09-10 04:34:40 +08:00
|
|
|
ArrayRef<PathSensitiveBugReport *> &bugReports);
|
2013-03-16 09:07:58 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
BugPathInfo *getNextBugPath();
|
2013-03-16 09:07:58 +08:00
|
|
|
};
|
2018-03-07 08:17:48 +08:00
|
|
|
|
|
|
|
} // namespace
|
2013-03-16 09:07:58 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
BugPathGetter::BugPathGetter(const ExplodedGraph *OriginalGraph,
|
2019-09-10 04:34:40 +08:00
|
|
|
ArrayRef<PathSensitiveBugReport *> &bugReports) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
SmallVector<const ExplodedNode *, 32> Nodes;
|
|
|
|
for (const auto I : bugReports) {
|
|
|
|
assert(I->isValid() &&
|
|
|
|
"We only allow BugReporterVisitors and BugReporter itself to "
|
|
|
|
"invalidate reports!");
|
|
|
|
Nodes.emplace_back(I->getErrorNode());
|
|
|
|
}
|
|
|
|
|
2013-03-20 08:35:37 +08:00
|
|
|
// The trimmed graph is created in the body of the constructor to ensure
|
|
|
|
// that the DenseMaps have been initialized already.
|
2013-03-23 05:15:28 +08:00
|
|
|
InterExplodedGraphMap ForwardMap;
|
2019-08-21 16:48:24 +08:00
|
|
|
TrimmedGraph = OriginalGraph->trim(Nodes, &ForwardMap);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Find the (first) error node in the trimmed graph. We just need to consult
|
2013-03-16 09:07:58 +08:00
|
|
|
// the node map which maps from nodes in the original graph to nodes
|
2009-04-01 07:00:32 +08:00
|
|
|
// in the new graph.
|
2013-03-23 05:15:28 +08:00
|
|
|
llvm::SmallPtrSet<const ExplodedNode *, 32> RemainingNodes;
|
2009-04-01 07:00:32 +08:00
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
for (PathSensitiveBugReport *Report : bugReports) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
const ExplodedNode *NewNode = ForwardMap.lookup(Report->getErrorNode());
|
|
|
|
assert(NewNode &&
|
|
|
|
"Failed to construct a trimmed graph that contains this error "
|
|
|
|
"node!");
|
|
|
|
ReportNodes.emplace_back(Report, NewNode);
|
|
|
|
RemainingNodes.insert(NewNode);
|
2010-12-03 14:52:30 +08:00
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
assert(!RemainingNodes.empty() && "No error node found in the trimmed graph");
|
|
|
|
|
|
|
|
// Perform a forward BFS to find all the shortest paths.
|
|
|
|
std::queue<const ExplodedNode *> WS;
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
assert(TrimmedGraph->num_roots() == 1);
|
|
|
|
WS.push(*TrimmedGraph->roots_begin());
|
2013-03-23 05:15:28 +08:00
|
|
|
unsigned Priority = 0;
|
2009-04-01 07:00:32 +08:00
|
|
|
|
|
|
|
while (!WS.empty()) {
|
2013-03-23 05:15:28 +08:00
|
|
|
const ExplodedNode *Node = WS.front();
|
2009-04-01 07:00:32 +08:00
|
|
|
WS.pop();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-20 08:35:37 +08:00
|
|
|
PriorityMapTy::iterator PriorityEntry;
|
|
|
|
bool IsNew;
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
std::tie(PriorityEntry, IsNew) = PriorityMap.insert({Node, Priority});
|
2013-03-23 05:15:28 +08:00
|
|
|
++Priority;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-20 08:35:37 +08:00
|
|
|
if (!IsNew) {
|
|
|
|
assert(PriorityEntry->second <= Priority);
|
|
|
|
continue;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
if (RemainingNodes.erase(Node))
|
|
|
|
if (RemainingNodes.empty())
|
|
|
|
break;
|
2013-03-19 07:34:37 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
for (const ExplodedNode *Succ : Node->succs())
|
|
|
|
WS.push(Succ);
|
2013-03-20 06:10:35 +08:00
|
|
|
}
|
2013-03-20 08:35:37 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
// Sort the error paths from longest to shortest.
|
2018-09-27 06:16:28 +08:00
|
|
|
llvm::sort(ReportNodes, PriorityCompare<true>(PriorityMap));
|
2013-03-23 05:15:28 +08:00
|
|
|
}
|
2013-03-20 08:35:37 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
BugPathInfo *BugPathGetter::getNextBugPath() {
|
2013-03-23 05:15:28 +08:00
|
|
|
if (ReportNodes.empty())
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
return nullptr;
|
2013-03-20 08:35:37 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
const ExplodedNode *OrigN;
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
std::tie(CurrentBugPath.Report, OrigN) = ReportNodes.pop_back_val();
|
2013-03-23 05:15:28 +08:00
|
|
|
assert(PriorityMap.find(OrigN) != PriorityMap.end() &&
|
|
|
|
"error node not accessible from root");
|
2013-03-20 08:35:37 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
// Create a new graph with a single path. This is the graph that will be
|
|
|
|
// returned to the caller.
|
2019-08-15 07:04:18 +08:00
|
|
|
auto GNew = std::make_unique<ExplodedGraph>();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
// Now walk from the error node up the BFS path, always taking the
|
|
|
|
// predeccessor with the lowest number.
|
2014-05-27 10:45:47 +08:00
|
|
|
ExplodedNode *Succ = nullptr;
|
2013-03-23 05:15:28 +08:00
|
|
|
while (true) {
|
2009-04-01 07:00:32 +08:00
|
|
|
// Create the equivalent node in the new graph with the same state
|
|
|
|
// and location.
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
ExplodedNode *NewN = GNew->createUncachedNode(
|
2019-10-18 07:10:09 +08:00
|
|
|
OrigN->getLocation(), OrigN->getState(),
|
|
|
|
OrigN->getID(), OrigN->isSink());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Link up the new node with the previous node.
|
2013-03-23 05:15:28 +08:00
|
|
|
if (Succ)
|
|
|
|
Succ->addPredecessor(NewN, *GNew);
|
|
|
|
else
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
CurrentBugPath.ErrorNode = NewN;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
Succ = NewN;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Are we at the final node?
|
2013-03-23 05:15:28 +08:00
|
|
|
if (OrigN->pred_empty()) {
|
|
|
|
GNew->addRoot(NewN);
|
2009-04-01 07:00:32 +08:00
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-03-23 05:15:28 +08:00
|
|
|
// Find the next predeccessor node. We choose the node that is marked
|
2013-03-20 08:35:37 +08:00
|
|
|
// with the lowest BFS number.
|
2013-03-23 05:15:28 +08:00
|
|
|
OrigN = *std::min_element(OrigN->pred_begin(), OrigN->pred_end(),
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
PriorityCompare<false>(PriorityMap));
|
2013-03-20 08:35:37 +08:00
|
|
|
}
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
CurrentBugPath.BugPath = std::move(GNew);
|
2014-09-05 08:11:25 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
return &CurrentBugPath;
|
2013-03-20 08:35:37 +08:00
|
|
|
}
|
|
|
|
|
2018-10-31 22:54:27 +08:00
|
|
|
/// CompactMacroExpandedPieces - This function postprocesses a PathDiagnostic
|
|
|
|
/// object and collapses PathDiagosticPieces that are expanded by macros.
|
|
|
|
static void CompactMacroExpandedPieces(PathPieces &path,
|
|
|
|
const SourceManager& SM) {
|
2019-08-14 00:45:48 +08:00
|
|
|
using MacroStackTy = std::vector<
|
|
|
|
std::pair<std::shared_ptr<PathDiagnosticMacroPiece>, SourceLocation>>;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2019-08-14 00:45:48 +08:00
|
|
|
using PiecesTy = std::vector<PathDiagnosticPieceRef>;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
MacroStackTy MacroStack;
|
|
|
|
PiecesTy Pieces;
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2013-04-30 07:12:59 +08:00
|
|
|
for (PathPieces::const_iterator I = path.begin(), E = path.end();
|
2018-03-07 08:17:48 +08:00
|
|
|
I != E; ++I) {
|
|
|
|
const auto &piece = *I;
|
2012-03-02 09:27:31 +08:00
|
|
|
|
|
|
|
// Recursively compact calls.
|
2017-01-06 01:26:53 +08:00
|
|
|
if (auto *call = dyn_cast<PathDiagnosticCallPiece>(&*piece)) {
|
2018-10-31 22:54:27 +08:00
|
|
|
CompactMacroExpandedPieces(call->path, SM);
|
2012-03-02 09:27:31 +08:00
|
|
|
}
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Get the location of the PathDiagnosticPiece.
|
2012-03-02 09:27:31 +08:00
|
|
|
const FullSourceLoc Loc = piece->getLocation().asLocation();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Determine the instantiation location, which is the location we group
|
|
|
|
// related PathDiagnosticPieces.
|
2009-09-09 23:08:12 +08:00
|
|
|
SourceLocation InstantiationLoc = Loc.isMacroID() ?
|
2011-07-26 00:49:02 +08:00
|
|
|
SM.getExpansionLoc(Loc) :
|
2009-04-01 07:00:32 +08:00
|
|
|
SourceLocation();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
if (Loc.isFileID()) {
|
|
|
|
MacroStack.clear();
|
2012-03-02 09:27:31 +08:00
|
|
|
Pieces.push_back(piece);
|
2009-04-01 07:00:32 +08:00
|
|
|
continue;
|
|
|
|
}
|
2008-04-08 07:35:17 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
assert(Loc.isMacroID());
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Is the PathDiagnosticPiece within the same macro group?
|
|
|
|
if (!MacroStack.empty() && InstantiationLoc == MacroStack.back().second) {
|
2012-03-02 09:27:31 +08:00
|
|
|
MacroStack.back().first->subPieces.push_back(piece);
|
2009-04-01 07:00:32 +08:00
|
|
|
continue;
|
|
|
|
}
|
2009-03-28 05:16:25 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// We aren't in the same group. Are we descending into a new macro
|
|
|
|
// or are part of an old one?
|
2017-01-06 01:26:53 +08:00
|
|
|
std::shared_ptr<PathDiagnosticMacroPiece> MacroGroup;
|
2009-03-28 05:16:25 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
SourceLocation ParentInstantiationLoc = InstantiationLoc.isMacroID() ?
|
2011-07-26 00:49:02 +08:00
|
|
|
SM.getExpansionLoc(Loc) :
|
2009-04-01 07:00:32 +08:00
|
|
|
SourceLocation();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Walk the entire macro stack.
|
|
|
|
while (!MacroStack.empty()) {
|
|
|
|
if (InstantiationLoc == MacroStack.back().second) {
|
|
|
|
MacroGroup = MacroStack.back().first;
|
|
|
|
break;
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
if (ParentInstantiationLoc == MacroStack.back().second) {
|
|
|
|
MacroGroup = MacroStack.back().first;
|
|
|
|
break;
|
2008-04-03 12:42:52 +08:00
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
MacroStack.pop_back();
|
2008-04-09 08:20:43 +08:00
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
if (!MacroGroup || ParentInstantiationLoc == MacroStack.back().second) {
|
|
|
|
// Create a new macro group and add it to the stack.
|
2017-01-06 01:26:53 +08:00
|
|
|
auto NewGroup = std::make_shared<PathDiagnosticMacroPiece>(
|
2012-03-02 09:27:31 +08:00
|
|
|
PathDiagnosticLocation::createSingleLocation(piece->getLocation()));
|
2008-04-24 07:35:07 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
if (MacroGroup)
|
2012-02-08 12:32:34 +08:00
|
|
|
MacroGroup->subPieces.push_back(NewGroup);
|
2009-04-01 07:00:32 +08:00
|
|
|
else {
|
|
|
|
assert(InstantiationLoc.isFileID());
|
|
|
|
Pieces.push_back(NewGroup);
|
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
MacroGroup = NewGroup;
|
|
|
|
MacroStack.push_back(std::make_pair(MacroGroup, InstantiationLoc));
|
2009-04-01 04:22:36 +08:00
|
|
|
}
|
2009-04-01 07:00:32 +08:00
|
|
|
|
|
|
|
// Finally, add the PathDiagnosticPiece to the group.
|
2012-03-02 09:27:31 +08:00
|
|
|
MacroGroup->subPieces.push_back(piece);
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-04-01 07:00:32 +08:00
|
|
|
// Now take the pieces and construct a new PathDiagnostic.
|
2012-03-02 09:27:31 +08:00
|
|
|
path.clear();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2015-02-18 00:48:30 +08:00
|
|
|
path.insert(path.end(), Pieces.begin(), Pieces.end());
|
2008-04-03 12:42:52 +08:00
|
|
|
}
|
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
/// Generate notes from all visitors.
|
2021-03-16 22:17:43 +08:00
|
|
|
/// Notes associated with @c ErrorNode are generated using
|
|
|
|
/// @c getEndPath, and the rest are generated with @c VisitNode.
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
static std::unique_ptr<VisitorsDiagnosticsTy>
|
2019-09-10 04:34:40 +08:00
|
|
|
generateVisitorsDiagnostics(PathSensitiveBugReport *R,
|
|
|
|
const ExplodedNode *ErrorNode,
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
BugReporterContext &BRC) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
std::unique_ptr<VisitorsDiagnosticsTy> Notes =
|
2019-08-15 07:04:18 +08:00
|
|
|
std::make_unique<VisitorsDiagnosticsTy>();
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport::VisitorList visitors;
|
2012-09-22 09:24:53 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Run visitors on all nodes starting from the node *before* the last one.
|
2021-03-16 22:17:43 +08:00
|
|
|
// The last node is reserved for notes generated with @c getEndPath.
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
const ExplodedNode *NextNode = ErrorNode->getFirstPred();
|
|
|
|
while (NextNode) {
|
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
// At each iteration, move all visitors from report to visitor list. This is
|
|
|
|
// important, because the Profile() functions of the visitors make sure that
|
|
|
|
// a visitor isn't added multiple times for the same node, but it's fine
|
|
|
|
// to add the a visitor with Profile() for different nodes (e.g. tracking
|
|
|
|
// a region at different points of the symbolic execution).
|
|
|
|
for (std::unique_ptr<BugReporterVisitor> &Visitor : R->visitors())
|
|
|
|
visitors.push_back(std::move(Visitor));
|
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
R->clearVisitors();
|
2009-09-09 23:08:12 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
const ExplodedNode *Pred = NextNode->getFirstPred();
|
|
|
|
if (!Pred) {
|
2019-08-14 00:45:48 +08:00
|
|
|
PathDiagnosticPieceRef LastPiece;
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
for (auto &V : visitors) {
|
|
|
|
V->finalizeVisitor(BRC, ErrorNode, *R);
|
2013-03-16 09:07:47 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
if (auto Piece = V->getEndPath(BRC, ErrorNode, *R)) {
|
|
|
|
assert(!LastPiece &&
|
|
|
|
"There can only be one final piece in a diagnostic.");
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
assert(Piece->getKind() == PathDiagnosticPiece::Kind::Event &&
|
|
|
|
"The final piece must contain a message!");
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
LastPiece = std::move(Piece);
|
|
|
|
(*Notes)[ErrorNode].push_back(LastPiece);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
2018-03-07 08:17:48 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
for (auto &V : visitors) {
|
2018-09-29 02:49:41 +08:00
|
|
|
auto P = V->VisitNode(NextNode, BRC, *R);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
if (P)
|
|
|
|
(*Notes)[NextNode].push_back(std::move(P));
|
|
|
|
}
|
2013-03-16 05:41:55 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
if (!R->isValid())
|
|
|
|
break;
|
|
|
|
|
|
|
|
NextNode = Pred;
|
|
|
|
}
|
|
|
|
|
|
|
|
return Notes;
|
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
Optional<PathDiagnosticBuilder> PathDiagnosticBuilder::findValidReport(
|
|
|
|
ArrayRef<PathSensitiveBugReport *> &bugReports,
|
|
|
|
PathSensitiveBugReporter &Reporter) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
|
|
|
|
BugPathGetter BugGraph(&Reporter.getGraph(), bugReports);
|
|
|
|
|
|
|
|
while (BugPathInfo *BugPath = BugGraph.getNextBugPath()) {
|
2013-03-16 05:41:55 +08:00
|
|
|
// Find the BugReport with the original location.
|
2019-09-10 04:34:40 +08:00
|
|
|
PathSensitiveBugReport *R = BugPath->Report;
|
2013-03-16 05:41:55 +08:00
|
|
|
assert(R && "No original report found for sliced graph.");
|
|
|
|
assert(R->isValid() && "Report selected by trimmed graph marked invalid.");
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
const ExplodedNode *ErrorNode = BugPath->ErrorNode;
|
2013-03-16 05:41:55 +08:00
|
|
|
|
[analyzer] False positive refutation with Z3
Summary: This is a prototype of a bug reporter visitor that invalidates bug reports by re-checking constraints of certain states on the bug path using the Z3 constraint manager backend. The functionality is available under the `crosscheck-with-z3` analyzer config flag.
Reviewers: george.karpenkov, NoQ, dcoughlin, rnkovacs
Reviewed By: george.karpenkov
Subscribers: rnkovacs, NoQ, george.karpenkov, dcoughlin, xbolva00, ddcc, mikhail.ramalho, MTC, fhahn, whisperity, baloghadamsoftware, szepet, a.sidorin, gsd, dkrupp, xazax.hun, cfe-commits
Differential Revision: https://reviews.llvm.org/D45517
llvm-svn: 333903
2018-06-04 22:40:44 +08:00
|
|
|
// Register refutation visitors first, if they mark the bug invalid no
|
|
|
|
// further analysis is required
|
2021-06-01 21:26:53 +08:00
|
|
|
R->addVisitor<LikelyFalsePositiveSuppressionBRVisitor>();
|
[analyzer] False positive refutation with Z3
Summary: This is a prototype of a bug reporter visitor that invalidates bug reports by re-checking constraints of certain states on the bug path using the Z3 constraint manager backend. The functionality is available under the `crosscheck-with-z3` analyzer config flag.
Reviewers: george.karpenkov, NoQ, dcoughlin, rnkovacs
Reviewed By: george.karpenkov
Subscribers: rnkovacs, NoQ, george.karpenkov, dcoughlin, xbolva00, ddcc, mikhail.ramalho, MTC, fhahn, whisperity, baloghadamsoftware, szepet, a.sidorin, gsd, dkrupp, xazax.hun, cfe-commits
Differential Revision: https://reviews.llvm.org/D45517
llvm-svn: 333903
2018-06-04 22:40:44 +08:00
|
|
|
|
2013-03-16 05:41:55 +08:00
|
|
|
// Register additional node visitors.
|
2021-06-01 21:26:53 +08:00
|
|
|
R->addVisitor<NilReceiverBRVisitor>();
|
|
|
|
R->addVisitor<ConditionBRVisitor>();
|
|
|
|
R->addVisitor<TagVisitor>();
|
2013-03-16 05:41:55 +08:00
|
|
|
|
2019-08-21 16:48:24 +08:00
|
|
|
BugReporterContext BRC(Reporter);
|
2013-01-31 03:12:34 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Run all visitors on a given graph, once.
|
|
|
|
std::unique_ptr<VisitorsDiagnosticsTy> visitorNotes =
|
|
|
|
generateVisitorsDiagnostics(R, ErrorNode, BRC);
|
2012-03-24 11:03:29 +08:00
|
|
|
|
2018-07-25 20:49:32 +08:00
|
|
|
if (R->isValid()) {
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
if (Reporter.getAnalyzerOptions().ShouldCrosscheckWithZ3) {
|
2018-07-25 20:49:32 +08:00
|
|
|
// If crosscheck is enabled, remove all visitors, add the refutation
|
|
|
|
// visitor and check again
|
|
|
|
R->clearVisitors();
|
2021-06-01 21:26:53 +08:00
|
|
|
R->addVisitor<FalsePositiveRefutationBRVisitor>();
|
2018-07-25 20:49:32 +08:00
|
|
|
|
2020-06-30 00:51:24 +08:00
|
|
|
// We don't overwrite the notes inserted by other visitors because the
|
2018-07-25 20:49:32 +08:00
|
|
|
// refutation manager does not add any new note to the path
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
generateVisitorsDiagnostics(R, BugPath->ErrorNode, BRC);
|
2018-07-25 20:49:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// Check if the bug is still valid
|
|
|
|
if (R->isValid())
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
return PathDiagnosticBuilder(
|
|
|
|
std::move(BRC), std::move(BugPath->BugPath), BugPath->Report,
|
|
|
|
BugPath->ErrorNode, std::move(visitorNotes));
|
2018-07-25 20:49:32 +08:00
|
|
|
}
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
2018-07-25 20:49:32 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P2.: Clean up the construction of bug paths and finding a valid report
This patch refactors the utility functions and classes around the construction
of a bug path.
At a very high level, this consists of 3 steps:
* For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
* Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
* Run all visitors on the constructed bug path. If in this process the report
got invalidated, start over from step 2.
Now, to the changes within this patch:
* Do not allow the invalidation of BugReports up to the point where the trimmed
graph is constructed. Checkers shouldn't add bug reports that are known to be
invalid, and should use visitors and argue about the entirety of the bug path if
needed.
* Do not calculate indices. I may be biased, but I personally find code like
this horrible. I'd like to point you to one of the comments in the original code:
SmallVector<const ExplodedNode *, 32> errorNodes;
for (const auto I : bugReports) {
if (I->isValid()) {
HasValid = true;
errorNodes.push_back(I->getErrorNode());
} else {
// Keep the errorNodes list in sync with the bugReports list.
errorNodes.push_back(nullptr);
}
}
Not on my watch. Instead, use a far easier to follow trick: store a pointer to
the BugReport in question, not an index to it.
* Add range iterators to ExplodedGraph's successors and predecessors, and a
visitor range to BugReporter.
* Rename TrimmedGraph to BugPathGetter. Because that is what it has always been:
no sane graph type should store an iterator-like state, or have an interface not
exposing a single graph-like functionalities.
* Rename ReportGraph to BugPathInfo, because it is only a linear path with some
other context.
* Instead of having both and out and in parameter (which I think isn't ever
excusable unless we use the out-param for caching), return a record object with
descriptive getter methods.
* Where descriptive names weren't sufficient, compliment the code with comments.
Differential Revision: https://reviews.llvm.org/D65379
llvm-svn: 368694
2019-08-13 21:56:12 +08:00
|
|
|
return {};
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
}
|
2013-03-16 05:41:55 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
std::unique_ptr<DiagnosticForConsumerMapTy>
|
2019-08-20 10:15:50 +08:00
|
|
|
PathSensitiveBugReporter::generatePathDiagnostics(
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
ArrayRef<PathDiagnosticConsumer *> consumers,
|
2019-09-10 04:34:40 +08:00
|
|
|
ArrayRef<PathSensitiveBugReport *> &bugReports) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
assert(!bugReports.empty());
|
2013-06-07 06:02:58 +08:00
|
|
|
|
2019-08-15 07:04:18 +08:00
|
|
|
auto Out = std::make_unique<DiagnosticForConsumerMapTy>();
|
2013-05-08 09:15:24 +08:00
|
|
|
|
[analyzer][NFC] Refactoring BugReporter.cpp P5.: Compact mile long function invocations into objects
In D65379, I briefly described the construction of bug paths from an
ExplodedGraph. This patch is about refactoring the code processing the bug path
into a bug report.
A part of finding a valid bug report was running all visitors on the bug path,
so we already have a (possibly empty) set of diagnostics for each ExplodedNode
in it.
Then, for each diagnostic consumer, we construct non-visitor diagnostic pieces.
* We first construct the final diagnostic piece (the warning), then
* We start ascending the bug path from the error node's predecessor (since the
error node itself was used to construct the warning event). For each node
* We check the location (whether its a CallEnter, CallExit) etc. We simultaneously
keep track of where we are with the execution by pushing CallStack when we see a
CallExit (keep in mind that everything is happening in reverse!), popping it
when we find a CallEnter, compacting them into a single PathDiagnosticCallEvent.
void f() {
bar();
}
void g() {
f();
error(); // warning
}
=== The bug path ===
(root) -> f's CallEnter -> bar() -> f's CallExit -> (error node)
=== Constructed report ===
f's CallEnter -> bar() -> f's CallExit
^ /
\ V
(root) ---> f's CallEvent --> (error node)
* We also keep track of different PathPieces different location contexts
* (CallEvent::path in the above example has f's LocationContext, while the
CallEvent itself is in g's context) in a LocationContextMap object. Construct
whatever piece, if any, is needed for the note.
* If we need to generate edges (or arrows) do so. Make sure to also connect
these pieces with the ones that visitors emitted.
* Clean up the constructed PathDiagnostic by making arrows nicer, pruning
function calls, etc.
So I complained about mile long function invocations with seemingly the same
parameters being passed around. This problem, as I see it, a natural candidate
for creating classes and tying them all together.
I tried very hard to make the implementation feel natural, like, rolling off the
tongue. I introduced 2 new classes: PathDiagnosticBuilder (I mean, I kept the
name but changed almost everything in it) contains every contextual information
(owns the bug path, the diagnostics constructed but the visitors, the BugReport
itself, etc) needed for constructing a PathDiagnostic object, and is pretty much
completely immutable. BugReportContruct is the object containing every
non-contextual information (the PathDiagnostic object we're constructing, the
current location in the bug path, the location context map and the call stack I
meantioned earlier), and is passed around all over the place as a single entity
instead of who knows how many parameters.
I tried to used constness, asserts, limiting visibility of fields to my
advantage to clean up the code big time and dramatically improve safety. Also,
whenever I found the code difficult to understand, I added comments and/or
examples.
Here's a complete list of changes and my design philosophy behind it:
* Instead of construcing a ReportInfo object (added by D65379) after finding a
valid bug report, simply return an optional PathDiagnosticBuilder object straight
away. Move findValidReport into the class as a static method. I find
GRBugReporter::generatePathDiagnostics a joy to look at now.
* Rename generatePathDiagnosticForConsumer to generate (maybe not needed, but
felt that way in the moment) and moved it to PathDiagnosticBuilder. If we don't
need to generate diagnostics, bail out straight away, like we always should have.
After that, construct a BugReportConstruct object, leaving the rest of the logic
untouched.
* Move all static methods that would use contextual information into
PathDiagnosticBuilder, reduce their parameter count drastically by simply
passing around a BugReportConstruct object.
* Glance at the code I removed: Could you tell what the original
PathDiagnosticBuilder::LC object was for? It took a gooood long while for me to
realize that nothing really. It is always equal with the LocationContext
associated with our current position in the bug path. Remove it completely.
* The original code contains the following expression quite a bit:
LCM[&PD.getActivePath()], so what does it mean? I said that we collect the
contexts associated with different PathPieces, but why would we ever modify that,
shouldn't it be set? Well, theoretically yes, but in the implementation, the
address of PathDiagnostic::getActivePath doesn't change if we move to an outer,
previously unexplored function. Add both descriptive method names and
explanations to BugReportConstruct to help on this.
* Add plenty of asserts, both for safety and as a poor man's documentation.
Differential Revision: https://reviews.llvm.org/D65484
llvm-svn: 368737
2019-08-14 03:01:33 +08:00
|
|
|
Optional<PathDiagnosticBuilder> PDB =
|
|
|
|
PathDiagnosticBuilder::findValidReport(bugReports, *this);
|
2013-06-06 08:12:37 +08:00
|
|
|
|
2019-08-16 09:53:14 +08:00
|
|
|
if (PDB) {
|
|
|
|
for (PathDiagnosticConsumer *PC : consumers) {
|
|
|
|
if (std::unique_ptr<PathDiagnostic> PD = PDB->generate(PC)) {
|
|
|
|
(*Out)[PC] = std::move(PD);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-09-22 09:24:53 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
return Out;
|
2009-04-01 07:00:32 +08:00
|
|
|
}
|
2008-05-23 07:45:19 +08:00
|
|
|
|
2015-06-23 21:15:32 +08:00
|
|
|
void BugReporter::emitReport(std::unique_ptr<BugReport> R) {
|
2019-09-10 04:34:40 +08:00
|
|
|
bool ValidSourceLoc = R->getLocation().isValid();
|
2013-08-02 06:16:30 +08:00
|
|
|
assert(ValidSourceLoc);
|
|
|
|
// If we mess up in a release build, we'd still prefer to just drop the bug
|
|
|
|
// instead of trying to go on.
|
|
|
|
if (!ValidSourceLoc)
|
|
|
|
return;
|
|
|
|
|
2009-02-05 07:49:09 +08:00
|
|
|
// Compute the bug report's hash to determine its equivalence class.
|
|
|
|
llvm::FoldingSetNodeID ID;
|
|
|
|
R->Profile(ID);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
|
|
|
// Lookup the equivance class. If there isn't one, create it.
|
2009-02-05 07:49:09 +08:00
|
|
|
void *InsertPos;
|
2011-02-23 08:16:01 +08:00
|
|
|
BugReportEquivClass* EQ = EQClasses.FindNodeOrInsertPos(ID, InsertPos);
|
2009-09-09 23:08:12 +08:00
|
|
|
|
2009-02-05 07:49:09 +08:00
|
|
|
if (!EQ) {
|
2015-06-23 21:15:32 +08:00
|
|
|
EQ = new BugReportEquivClass(std::move(R));
|
2011-02-23 08:16:01 +08:00
|
|
|
EQClasses.InsertNode(EQ, InsertPos);
|
2011-08-19 09:57:09 +08:00
|
|
|
EQClassesVector.push_back(EQ);
|
2014-08-30 03:57:52 +08:00
|
|
|
} else
|
2015-06-23 21:15:32 +08:00
|
|
|
EQ->AddReport(std::move(R));
|
2008-04-03 12:42:52 +08:00
|
|
|
}
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
void PathSensitiveBugReporter::emitReport(std::unique_ptr<BugReport> R) {
|
|
|
|
if (auto PR = dyn_cast<PathSensitiveBugReport>(R.get()))
|
|
|
|
if (const ExplodedNode *E = PR->getErrorNode()) {
|
|
|
|
// An error node must either be a sink or have a tag, otherwise
|
|
|
|
// it could get reclaimed before the path diagnostic is created.
|
|
|
|
assert((E->isSink() || E->getLocation().getTag()) &&
|
|
|
|
"Error node must either be a sink or have a tag");
|
|
|
|
|
|
|
|
const AnalysisDeclContext *DeclCtx =
|
|
|
|
E->getLocationContext()->getAnalysisDeclContext();
|
|
|
|
// The source of autosynthesized body can be handcrafted AST or a model
|
|
|
|
// file. The locations from handcrafted ASTs have no valid source
|
|
|
|
// locations and have to be discarded. Locations from model files should
|
|
|
|
// be preserved for processing and reporting.
|
|
|
|
if (DeclCtx->isBodyAutosynthesized() &&
|
|
|
|
!DeclCtx->isBodyAutosynthesizedFromModelFile())
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
BugReporter::emitReport(std::move(R));
|
|
|
|
}
|
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Emitting reports in equivalence classes.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
2018-03-07 08:17:48 +08:00
|
|
|
|
2009-11-28 14:07:30 +08:00
|
|
|
struct FRIEC_WLItem {
|
2009-09-15 06:01:32 +08:00
|
|
|
const ExplodedNode *N;
|
|
|
|
ExplodedNode::const_succ_iterator I, E;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
FRIEC_WLItem(const ExplodedNode *n)
|
2018-03-07 08:17:48 +08:00
|
|
|
: N(n), I(N->succ_begin()), E(N->succ_end()) {}
|
2015-09-08 11:50:52 +08:00
|
|
|
};
|
2018-03-07 08:17:48 +08:00
|
|
|
|
|
|
|
} // namespace
|
2009-09-15 06:01:32 +08:00
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
BugReport *PathSensitiveBugReporter::findReportInEquivalenceClass(
|
|
|
|
BugReportEquivClass &EQ, SmallVectorImpl<BugReport *> &bugReports) {
|
2010-12-03 14:52:30 +08:00
|
|
|
// If we don't need to suppress any of the nodes because they are
|
|
|
|
// post-dominated by a sink, simply add all the nodes in the equivalence class
|
|
|
|
// to 'Nodes'. Any of the reports will serve as a "representative" report.
|
2019-09-10 04:34:44 +08:00
|
|
|
assert(EQ.getReports().size() > 0);
|
|
|
|
const BugType& BT = EQ.getReports()[0]->getBugType();
|
2010-09-10 03:05:34 +08:00
|
|
|
if (!BT.isSuppressOnSink()) {
|
2019-09-10 04:34:44 +08:00
|
|
|
BugReport *R = EQ.getReports()[0].get();
|
|
|
|
for (auto &J : EQ.getReports()) {
|
|
|
|
if (auto *PR = dyn_cast<PathSensitiveBugReport>(J.get())) {
|
2019-09-10 04:34:40 +08:00
|
|
|
R = PR;
|
|
|
|
bugReports.push_back(PR);
|
2010-09-10 03:05:34 +08:00
|
|
|
}
|
|
|
|
}
|
2009-09-15 06:01:32 +08:00
|
|
|
return R;
|
2010-09-10 03:05:34 +08:00
|
|
|
}
|
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
// For bug reports that should be suppressed when all paths are post-dominated
|
|
|
|
// by a sink node, iterate through the reports in the equivalence class
|
|
|
|
// until we find one that isn't post-dominated (if one exists). We use a
|
|
|
|
// DFS traversal of the ExplodedGraph to find a non-sink node. We could write
|
|
|
|
// this as a recursive function, but we don't want to risk blowing out the
|
|
|
|
// stack for very long paths.
|
2014-05-27 10:45:47 +08:00
|
|
|
BugReport *exampleReport = nullptr;
|
2010-09-10 03:05:34 +08:00
|
|
|
|
2019-09-10 04:34:44 +08:00
|
|
|
for (const auto &I: EQ.getReports()) {
|
|
|
|
auto *R = dyn_cast<PathSensitiveBugReport>(I.get());
|
2019-09-10 04:34:40 +08:00
|
|
|
if (!R)
|
2009-09-15 06:01:32 +08:00
|
|
|
continue;
|
2019-09-10 04:34:40 +08:00
|
|
|
|
|
|
|
const ExplodedNode *errorNode = R->getErrorNode();
|
2010-12-03 14:52:30 +08:00
|
|
|
if (errorNode->isSink()) {
|
2011-09-23 13:06:16 +08:00
|
|
|
llvm_unreachable(
|
2009-09-15 06:01:32 +08:00
|
|
|
"BugType::isSuppressSink() should not be 'true' for sink end nodes");
|
|
|
|
}
|
2010-09-10 03:05:34 +08:00
|
|
|
// No successors? By definition this nodes isn't post-dominated by a sink.
|
2010-12-03 14:52:30 +08:00
|
|
|
if (errorNode->succ_empty()) {
|
2019-09-10 04:34:40 +08:00
|
|
|
bugReports.push_back(R);
|
2010-12-03 14:52:30 +08:00
|
|
|
if (!exampleReport)
|
2019-09-10 04:34:40 +08:00
|
|
|
exampleReport = R;
|
2010-09-10 03:05:34 +08:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2016-12-22 22:48:52 +08:00
|
|
|
// See if we are in a no-return CFG block. If so, treat this similarly
|
|
|
|
// to being post-dominated by a sink. This works better when the analysis
|
2017-08-14 16:38:47 +08:00
|
|
|
// is incomplete and we have never reached the no-return function call(s)
|
|
|
|
// that we'd inevitably bump into on this path.
|
2019-08-14 20:20:08 +08:00
|
|
|
if (const CFGBlock *ErrorB = errorNode->getCFGBlock())
|
|
|
|
if (ErrorB->isInevitablySinking())
|
|
|
|
continue;
|
2016-12-22 22:48:52 +08:00
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
// At this point we know that 'N' is not a sink and it has at least one
|
2015-09-08 11:50:52 +08:00
|
|
|
// successor. Use a DFS worklist to find a non-sink end-of-path node.
|
2018-03-07 08:17:48 +08:00
|
|
|
using WLItem = FRIEC_WLItem;
|
|
|
|
using DFSWorkList = SmallVector<WLItem, 10>;
|
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
llvm::DenseMap<const ExplodedNode *, unsigned> Visited;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
DFSWorkList WL;
|
2010-12-03 14:52:30 +08:00
|
|
|
WL.push_back(errorNode);
|
|
|
|
Visited[errorNode] = 1;
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
while (!WL.empty()) {
|
|
|
|
WLItem &WI = WL.back();
|
|
|
|
assert(!WI.N->succ_empty());
|
2015-09-08 11:50:52 +08:00
|
|
|
|
2009-09-15 06:01:32 +08:00
|
|
|
for (; WI.I != WI.E; ++WI.I) {
|
2015-09-08 11:50:52 +08:00
|
|
|
const ExplodedNode *Succ = *WI.I;
|
2009-09-15 06:01:32 +08:00
|
|
|
// End-of-path node?
|
|
|
|
if (Succ->succ_empty()) {
|
2010-09-10 03:05:34 +08:00
|
|
|
// If we found an end-of-path node that is not a sink.
|
|
|
|
if (!Succ->isSink()) {
|
2019-09-10 04:34:40 +08:00
|
|
|
bugReports.push_back(R);
|
2010-12-03 14:52:30 +08:00
|
|
|
if (!exampleReport)
|
2019-09-10 04:34:40 +08:00
|
|
|
exampleReport = R;
|
2010-09-10 03:05:34 +08:00
|
|
|
WL.clear();
|
|
|
|
break;
|
|
|
|
}
|
2009-09-15 06:01:32 +08:00
|
|
|
// Found a sink? Continue on to the next successor.
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
// Mark the successor as visited. If it hasn't been explored,
|
|
|
|
// enqueue it to the DFS worklist.
|
|
|
|
unsigned &mark = Visited[Succ];
|
|
|
|
if (!mark) {
|
|
|
|
mark = 1;
|
|
|
|
WL.push_back(Succ);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2010-09-10 03:05:34 +08:00
|
|
|
|
|
|
|
// The worklist may have been cleared at this point. First
|
|
|
|
// check if it is empty before checking the last item.
|
|
|
|
if (!WL.empty() && &WL.back() == &WI)
|
2009-09-15 06:01:32 +08:00
|
|
|
WL.pop_back();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2010-09-10 03:05:34 +08:00
|
|
|
// ExampleReport will be NULL if all the nodes in the equivalence class
|
|
|
|
// were post-dominated by sinks.
|
2010-12-03 14:52:30 +08:00
|
|
|
return exampleReport;
|
2010-09-10 03:05:34 +08:00
|
|
|
}
|
2009-09-19 06:37:37 +08:00
|
|
|
|
2009-02-05 07:49:09 +08:00
|
|
|
void BugReporter::FlushReport(BugReportEquivClass& EQ) {
|
2011-07-23 18:55:15 +08:00
|
|
|
SmallVector<BugReport*, 10> bugReports;
|
2019-09-10 04:34:40 +08:00
|
|
|
BugReport *report = findReportInEquivalenceClass(EQ, bugReports);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
if (!report)
|
|
|
|
return;
|
|
|
|
|
[analyzer] Make checker silencing work for non-pathsensitive bug reports
D66572 separated BugReport and BugReporter into basic and path sensitive
versions. As a result, checker silencing, which worked deep in the path
sensitive report generation facilities became specific to it. DeadStoresChecker,
for instance, despite being in the static analyzer, emits non-pathsensitive
reports, and was impossible to silence.
This patch moves the corresponding code before the call to the virtual function
generateDiagnosticForConsumerMap (which is overriden by the specific kinds of
bug reporters). Although we see bug reporting as relatively lightweight compared
to the analysis, this will get rid of several steps we used to throw away.
Quoting from D65379:
At a very high level, this consists of 3 steps:
For all BugReports in the same BugReportEquivClass, collect all their error
nodes in a set. With that set, create a new, trimmed ExplodedGraph whose leafs
are all error nodes.
Until a valid report is found, construct a bug path, which is yet another
ExplodedGraph, that is linear from a given error node to the root of the graph.
Run all visitors on the constructed bug path. If in this process the report got
invalidated, start over from step 2.
Checker silencing used to kick in after all of these. Now it does before any of
them :^)
Differential Revision: https://reviews.llvm.org/D102914
Change-Id: Ice42939304516f2bebd05a1ea19878b89c96a25d
2021-05-21 20:02:03 +08:00
|
|
|
// See whether we need to silence the checker/package.
|
|
|
|
for (const std::string &CheckerOrPackage :
|
|
|
|
getAnalyzerOptions().SilencedCheckersAndPackages) {
|
|
|
|
if (report->getBugType().getCheckerName().startswith(
|
|
|
|
CheckerOrPackage))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
ArrayRef<PathDiagnosticConsumer*> Consumers = getPathDiagnosticConsumers();
|
|
|
|
std::unique_ptr<DiagnosticForConsumerMapTy> Diagnostics =
|
|
|
|
generateDiagnosticForConsumerMap(report, Consumers, bugReports);
|
|
|
|
|
|
|
|
for (auto &P : *Diagnostics) {
|
|
|
|
PathDiagnosticConsumer *Consumer = P.first;
|
|
|
|
std::unique_ptr<PathDiagnostic> &PD = P.second;
|
|
|
|
|
|
|
|
// If the path is empty, generate a single step path with the location
|
|
|
|
// of the issue.
|
|
|
|
if (PD->path.empty()) {
|
2019-09-10 04:34:40 +08:00
|
|
|
PathDiagnosticLocation L = report->getLocation();
|
2019-08-15 07:04:18 +08:00
|
|
|
auto piece = std::make_unique<PathDiagnosticEventPiece>(
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
L, report->getDescription());
|
|
|
|
for (SourceRange Range : report->getRanges())
|
|
|
|
piece->addRange(Range);
|
|
|
|
PD->setEndOfPath(std::move(piece));
|
|
|
|
}
|
|
|
|
|
|
|
|
PathPieces &Pieces = PD->getMutablePieces();
|
[analyzer] Evaluate all non-checker config options before analysis
In earlier patches regarding AnalyzerOptions, a lot of effort went into
gathering all config options, and changing the interface so that potential
misuse can be eliminited.
Up until this point, AnalyzerOptions only evaluated an option when it was
querried. For example, if we had a "-no-false-positives" flag, AnalyzerOptions
would store an Optional field for it that would be None up until somewhere in
the code until the flag's getter function is called.
However, now that we're confident that we've gathered all configs, we can
evaluate off of them before analysis, so we can emit a error on invalid input
even if that prticular flag will not matter in that particular run of the
analyzer. Another very big benefit of this is that debug.ConfigDumper will now
show the value of all configs every single time.
Also, almost all options related class have a similar interface, so uniformity
is also a benefit.
The implementation for errors on invalid input will be commited shorty.
Differential Revision: https://reviews.llvm.org/D53692
llvm-svn: 348031
2018-12-01 04:44:00 +08:00
|
|
|
if (getAnalyzerOptions().ShouldDisplayNotesAsEvents) {
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// For path diagnostic consumers that don't support extra notes,
|
|
|
|
// we may optionally convert those to path notes.
|
|
|
|
for (auto I = report->getNotes().rbegin(),
|
|
|
|
E = report->getNotes().rend(); I != E; ++I) {
|
|
|
|
PathDiagnosticNotePiece *Piece = I->get();
|
|
|
|
auto ConvertedPiece = std::make_shared<PathDiagnosticEventPiece>(
|
|
|
|
Piece->getLocation(), Piece->getString());
|
|
|
|
for (const auto &R: Piece->getRanges())
|
|
|
|
ConvertedPiece->addRange(R);
|
|
|
|
|
|
|
|
Pieces.push_front(std::move(ConvertedPiece));
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
for (auto I = report->getNotes().rbegin(),
|
|
|
|
E = report->getNotes().rend(); I != E; ++I)
|
|
|
|
Pieces.push_front(*I);
|
Allow multiple PathDiagnosticConsumers to be used with a BugReporter at the same time.
This fixes several issues:
- removes egregious hack where PlistDiagnosticConsumer would forward to HTMLDiagnosticConsumer,
but diagnostics wouldn't be generated consistently in the same way if PlistDiagnosticConsumer
was used by itself.
- emitting diagnostics to the terminal (using clang's diagnostic machinery) is no longer a special
case, just another PathDiagnosticConsumer. This also magically resolved some duplicate warnings,
as we now use PathDiagnosticConsumer's diagnostic pruning, which has scope for the entire translation
unit, not just the scope of a BugReporter (which is limited to a particular ExprEngine).
As an interesting side-effect, diagnostics emitted to the terminal also have their trailing "." stripped,
just like with diagnostics emitted to plists and HTML. This required some tests to be updated, but now
the tests have higher fidelity with what users will see.
There are some inefficiencies in this patch. We currently generate the report graph (from the ExplodedGraph)
once per PathDiagnosticConsumer, which is a bit wasteful, but that could be pulled up higher in the
logic stack. There is some intended duplication, however, as we now generate different PathDiagnostics (for the same issue)
for different PathDiagnosticConsumers. This is necessary to produce the diagnostics that a particular
consumer expects.
llvm-svn: 162028
2012-08-17 01:45:23 +08:00
|
|
|
}
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
|
2019-09-07 04:55:29 +08:00
|
|
|
for (const auto &I : report->getFixits())
|
|
|
|
Pieces.back()->addFixit(I);
|
|
|
|
|
2018-09-07 08:42:53 +08:00
|
|
|
updateExecutedLinesWithDiagnosticPieces(*PD);
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
Consumer->HandlePathDiagnostic(std::move(PD));
|
Allow multiple PathDiagnosticConsumers to be used with a BugReporter at the same time.
This fixes several issues:
- removes egregious hack where PlistDiagnosticConsumer would forward to HTMLDiagnosticConsumer,
but diagnostics wouldn't be generated consistently in the same way if PlistDiagnosticConsumer
was used by itself.
- emitting diagnostics to the terminal (using clang's diagnostic machinery) is no longer a special
case, just another PathDiagnosticConsumer. This also magically resolved some duplicate warnings,
as we now use PathDiagnosticConsumer's diagnostic pruning, which has scope for the entire translation
unit, not just the scope of a BugReporter (which is limited to a particular ExprEngine).
As an interesting side-effect, diagnostics emitted to the terminal also have their trailing "." stripped,
just like with diagnostics emitted to plists and HTML. This required some tests to be updated, but now
the tests have higher fidelity with what users will see.
There are some inefficiencies in this patch. We currently generate the report graph (from the ExplodedGraph)
once per PathDiagnosticConsumer, which is a bit wasteful, but that could be pulled up higher in the
logic stack. There is some intended duplication, however, as we now generate different PathDiagnostics (for the same issue)
for different PathDiagnosticConsumers. This is necessary to produce the diagnostics that a particular
consumer expects.
llvm-svn: 162028
2012-08-17 01:45:23 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-01-17 10:59:11 +08:00
|
|
|
/// Insert all lines participating in the function signature \p Signature
|
|
|
|
/// into \p ExecutedLines.
|
|
|
|
static void populateExecutedLinesWithFunctionSignature(
|
2019-08-14 02:48:08 +08:00
|
|
|
const Decl *Signature, const SourceManager &SM,
|
2018-09-07 08:43:55 +08:00
|
|
|
FilesToLineNumsMap &ExecutedLines) {
|
2018-01-17 10:59:11 +08:00
|
|
|
SourceRange SignatureSourceRange;
|
|
|
|
const Stmt* Body = Signature->getBody();
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto FD = dyn_cast<FunctionDecl>(Signature)) {
|
2018-01-17 10:59:11 +08:00
|
|
|
SignatureSourceRange = FD->getSourceRange();
|
2018-03-07 08:17:48 +08:00
|
|
|
} else if (const auto OD = dyn_cast<ObjCMethodDecl>(Signature)) {
|
2018-01-17 10:59:11 +08:00
|
|
|
SignatureSourceRange = OD->getSourceRange();
|
|
|
|
} else {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
SourceLocation Start = SignatureSourceRange.getBegin();
|
|
|
|
SourceLocation End = Body ? Body->getSourceRange().getBegin()
|
|
|
|
: SignatureSourceRange.getEnd();
|
2018-09-07 08:43:17 +08:00
|
|
|
if (!Start.isValid() || !End.isValid())
|
|
|
|
return;
|
2018-01-17 10:59:11 +08:00
|
|
|
unsigned StartLine = SM.getExpansionLineNumber(Start);
|
|
|
|
unsigned EndLine = SM.getExpansionLineNumber(End);
|
|
|
|
|
|
|
|
FileID FID = SM.getFileID(SM.getExpansionLoc(Start));
|
|
|
|
for (unsigned Line = StartLine; Line <= EndLine; Line++)
|
2018-09-07 08:43:55 +08:00
|
|
|
ExecutedLines[FID].insert(Line);
|
2018-01-17 10:59:11 +08:00
|
|
|
}
|
|
|
|
|
2018-02-24 07:26:54 +08:00
|
|
|
static void populateExecutedLinesWithStmt(
|
2019-08-14 02:48:08 +08:00
|
|
|
const Stmt *S, const SourceManager &SM,
|
2018-09-07 08:43:55 +08:00
|
|
|
FilesToLineNumsMap &ExecutedLines) {
|
2018-02-24 07:26:54 +08:00
|
|
|
SourceLocation Loc = S->getSourceRange().getBegin();
|
2018-09-07 08:43:17 +08:00
|
|
|
if (!Loc.isValid())
|
|
|
|
return;
|
2018-02-24 07:26:54 +08:00
|
|
|
SourceLocation ExpansionLoc = SM.getExpansionLoc(Loc);
|
|
|
|
FileID FID = SM.getFileID(ExpansionLoc);
|
|
|
|
unsigned LineNo = SM.getExpansionLineNumber(ExpansionLoc);
|
2018-09-07 08:43:55 +08:00
|
|
|
ExecutedLines[FID].insert(LineNo);
|
2018-02-24 07:26:54 +08:00
|
|
|
}
|
|
|
|
|
2018-01-17 10:59:11 +08:00
|
|
|
/// \return all executed lines including function signatures on the path
|
|
|
|
/// starting from \p N.
|
|
|
|
static std::unique_ptr<FilesToLineNumsMap>
|
2019-08-14 02:48:08 +08:00
|
|
|
findExecutedLines(const SourceManager &SM, const ExplodedNode *N) {
|
2019-08-15 07:04:18 +08:00
|
|
|
auto ExecutedLines = std::make_unique<FilesToLineNumsMap>();
|
2018-01-17 10:59:11 +08:00
|
|
|
|
|
|
|
while (N) {
|
|
|
|
if (N->getFirstPred() == nullptr) {
|
|
|
|
// First node: show signature of the entrance point.
|
|
|
|
const Decl *D = N->getLocationContext()->getDecl();
|
2018-09-07 08:43:55 +08:00
|
|
|
populateExecutedLinesWithFunctionSignature(D, SM, *ExecutedLines);
|
2018-01-17 10:59:11 +08:00
|
|
|
} else if (auto CE = N->getLocationAs<CallEnter>()) {
|
|
|
|
// Inlined function: show signature.
|
|
|
|
const Decl* D = CE->getCalleeContext()->getDecl();
|
2018-09-07 08:43:55 +08:00
|
|
|
populateExecutedLinesWithFunctionSignature(D, SM, *ExecutedLines);
|
2019-09-12 04:54:21 +08:00
|
|
|
} else if (const Stmt *S = N->getStmtForDiagnostics()) {
|
2018-09-07 08:43:55 +08:00
|
|
|
populateExecutedLinesWithStmt(S, SM, *ExecutedLines);
|
2018-02-24 07:26:54 +08:00
|
|
|
|
|
|
|
// Show extra context for some parent kinds.
|
|
|
|
const Stmt *P = N->getParentMap().getParent(S);
|
|
|
|
|
|
|
|
// The path exploration can die before the node with the associated
|
|
|
|
// return statement is generated, but we do want to show the whole
|
|
|
|
// return.
|
2018-03-07 08:17:48 +08:00
|
|
|
if (const auto *RS = dyn_cast_or_null<ReturnStmt>(P)) {
|
2018-09-07 08:43:55 +08:00
|
|
|
populateExecutedLinesWithStmt(RS, SM, *ExecutedLines);
|
2018-02-24 07:26:54 +08:00
|
|
|
P = N->getParentMap().getParent(RS);
|
|
|
|
}
|
2018-01-17 10:59:11 +08:00
|
|
|
|
2021-10-20 23:43:31 +08:00
|
|
|
if (isa_and_nonnull<SwitchCase, LabelStmt>(P))
|
2018-09-07 08:43:55 +08:00
|
|
|
populateExecutedLinesWithStmt(P, SM, *ExecutedLines);
|
2018-01-17 10:59:11 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
N = N->getFirstPred();
|
|
|
|
}
|
|
|
|
return ExecutedLines;
|
|
|
|
}
|
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
std::unique_ptr<DiagnosticForConsumerMapTy>
|
|
|
|
BugReporter::generateDiagnosticForConsumerMap(
|
2019-09-10 04:34:40 +08:00
|
|
|
BugReport *exampleReport, ArrayRef<PathDiagnosticConsumer *> consumers,
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
ArrayRef<BugReport *> bugReports) {
|
2019-09-10 04:34:40 +08:00
|
|
|
auto *basicReport = cast<BasicBugReport>(exampleReport);
|
|
|
|
auto Out = std::make_unique<DiagnosticForConsumerMapTy>();
|
|
|
|
for (auto *Consumer : consumers)
|
|
|
|
(*Out)[Consumer] = generateDiagnosticForBasicReport(basicReport);
|
|
|
|
return Out;
|
|
|
|
}
|
2009-04-30 05:58:13 +08:00
|
|
|
|
2019-09-12 04:54:24 +08:00
|
|
|
static PathDiagnosticCallPiece *
|
|
|
|
getFirstStackedCallToHeaderFile(PathDiagnosticCallPiece *CP,
|
|
|
|
const SourceManager &SMgr) {
|
|
|
|
SourceLocation CallLoc = CP->callEnter.asLocation();
|
|
|
|
|
|
|
|
// If the call is within a macro, don't do anything (for now).
|
|
|
|
if (CallLoc.isMacroID())
|
|
|
|
return nullptr;
|
|
|
|
|
|
|
|
assert(AnalysisManager::isInCodeFile(CallLoc, SMgr) &&
|
|
|
|
"The call piece should not be in a header file.");
|
|
|
|
|
|
|
|
// Check if CP represents a path through a function outside of the main file.
|
|
|
|
if (!AnalysisManager::isInCodeFile(CP->callEnterWithin.asLocation(), SMgr))
|
|
|
|
return CP;
|
|
|
|
|
|
|
|
const PathPieces &Path = CP->path;
|
|
|
|
if (Path.empty())
|
|
|
|
return nullptr;
|
|
|
|
|
|
|
|
// Check if the last piece in the callee path is a call to a function outside
|
|
|
|
// of the main file.
|
|
|
|
if (auto *CPInner = dyn_cast<PathDiagnosticCallPiece>(Path.back().get()))
|
|
|
|
return getFirstStackedCallToHeaderFile(CPInner, SMgr);
|
|
|
|
|
|
|
|
// Otherwise, the last piece is in the main file.
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void resetDiagnosticLocationToMainFile(PathDiagnostic &PD) {
|
|
|
|
if (PD.path.empty())
|
|
|
|
return;
|
|
|
|
|
|
|
|
PathDiagnosticPiece *LastP = PD.path.back().get();
|
|
|
|
assert(LastP);
|
|
|
|
const SourceManager &SMgr = LastP->getLocation().getManager();
|
|
|
|
|
|
|
|
// We only need to check if the report ends inside headers, if the last piece
|
|
|
|
// is a call piece.
|
|
|
|
if (auto *CP = dyn_cast<PathDiagnosticCallPiece>(LastP)) {
|
|
|
|
CP = getFirstStackedCallToHeaderFile(CP, SMgr);
|
|
|
|
if (CP) {
|
|
|
|
// Mark the piece.
|
|
|
|
CP->setAsLastInMainSourceFile();
|
|
|
|
|
|
|
|
// Update the path diagnostic message.
|
|
|
|
const auto *ND = dyn_cast<NamedDecl>(CP->getCallee());
|
|
|
|
if (ND) {
|
|
|
|
SmallString<200> buf;
|
|
|
|
llvm::raw_svector_ostream os(buf);
|
|
|
|
os << " (within a call to '" << ND->getDeclName() << "')";
|
|
|
|
PD.appendToDesc(os.str());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Reset the report containing declaration and location.
|
|
|
|
PD.setDeclWithIssue(CP->getCaller());
|
|
|
|
PD.setLocation(CP->getLocation());
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-09-10 04:34:40 +08:00
|
|
|
std::unique_ptr<DiagnosticForConsumerMapTy>
|
|
|
|
PathSensitiveBugReporter::generateDiagnosticForConsumerMap(
|
|
|
|
BugReport *exampleReport, ArrayRef<PathDiagnosticConsumer *> consumers,
|
|
|
|
ArrayRef<BugReport *> bugReports) {
|
|
|
|
std::vector<BasicBugReport *> BasicBugReports;
|
|
|
|
std::vector<PathSensitiveBugReport *> PathSensitiveBugReports;
|
|
|
|
if (isa<BasicBugReport>(exampleReport))
|
|
|
|
return BugReporter::generateDiagnosticForConsumerMap(exampleReport,
|
|
|
|
consumers, bugReports);
|
2016-10-08 03:25:10 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Generate the full path sensitive diagnostic, using the generation scheme
|
|
|
|
// specified by the PathDiagnosticConsumer. Note that we have to generate
|
|
|
|
// path diagnostics even for consumers which do not support paths, because
|
|
|
|
// the BugReporterVisitors may mark this bug as a false positive.
|
|
|
|
assert(!bugReports.empty());
|
|
|
|
MaxBugClassSize.updateMax(bugReports.size());
|
2019-09-10 04:34:40 +08:00
|
|
|
|
|
|
|
// Avoid copying the whole array because there may be a lot of reports.
|
|
|
|
ArrayRef<PathSensitiveBugReport *> convertedArrayOfReports(
|
|
|
|
reinterpret_cast<PathSensitiveBugReport *const *>(&*bugReports.begin()),
|
|
|
|
reinterpret_cast<PathSensitiveBugReport *const *>(&*bugReports.end()));
|
|
|
|
std::unique_ptr<DiagnosticForConsumerMapTy> Out = generatePathDiagnostics(
|
|
|
|
consumers, convertedArrayOfReports);
|
2016-10-08 03:25:10 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
if (Out->empty())
|
|
|
|
return Out;
|
2009-01-24 08:55:43 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
MaxValidBugClassSize.updateMax(bugReports.size());
|
2013-03-16 05:41:53 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
// Examine the report and see if the last piece is in a header. Reset the
|
|
|
|
// report location to the last piece in the main source file.
|
2019-08-14 02:48:08 +08:00
|
|
|
const AnalyzerOptions &Opts = getAnalyzerOptions();
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
for (auto const &P : *Out)
|
[analyzer] Evaluate all non-checker config options before analysis
In earlier patches regarding AnalyzerOptions, a lot of effort went into
gathering all config options, and changing the interface so that potential
misuse can be eliminited.
Up until this point, AnalyzerOptions only evaluated an option when it was
querried. For example, if we had a "-no-false-positives" flag, AnalyzerOptions
would store an Optional field for it that would be None up until somewhere in
the code until the flag's getter function is called.
However, now that we're confident that we've gathered all configs, we can
evaluate off of them before analysis, so we can emit a error on invalid input
even if that prticular flag will not matter in that particular run of the
analyzer. Another very big benefit of this is that debug.ConfigDumper will now
show the value of all configs every single time.
Also, almost all options related class have a similar interface, so uniformity
is also a benefit.
The implementation for errors on invalid input will be commited shorty.
Differential Revision: https://reviews.llvm.org/D53692
llvm-svn: 348031
2018-12-01 04:44:00 +08:00
|
|
|
if (Opts.ShouldReportIssuesInMainSourceFile && !Opts.AnalyzeAll)
|
2019-09-12 04:54:24 +08:00
|
|
|
resetDiagnosticLocationToMainFile(*P.second);
|
Allow multiple PathDiagnosticConsumers to be used with a BugReporter at the same time.
This fixes several issues:
- removes egregious hack where PlistDiagnosticConsumer would forward to HTMLDiagnosticConsumer,
but diagnostics wouldn't be generated consistently in the same way if PlistDiagnosticConsumer
was used by itself.
- emitting diagnostics to the terminal (using clang's diagnostic machinery) is no longer a special
case, just another PathDiagnosticConsumer. This also magically resolved some duplicate warnings,
as we now use PathDiagnosticConsumer's diagnostic pruning, which has scope for the entire translation
unit, not just the scope of a BugReporter (which is limited to a particular ExprEngine).
As an interesting side-effect, diagnostics emitted to the terminal also have their trailing "." stripped,
just like with diagnostics emitted to plists and HTML. This required some tests to be updated, but now
the tests have higher fidelity with what users will see.
There are some inefficiencies in this patch. We currently generate the report graph (from the ExplodedGraph)
once per PathDiagnosticConsumer, which is a bit wasteful, but that could be pulled up higher in the
logic stack. There is some intended duplication, however, as we now generate different PathDiagnostics (for the same issue)
for different PathDiagnosticConsumers. This is necessary to produce the diagnostics that a particular
consumer expects.
llvm-svn: 162028
2012-08-17 01:45:23 +08:00
|
|
|
|
[analyzer] Do not run visitors until the fixpoint, run only once.
In the current implementation, we run visitors until the fixed point is
reached.
That is, if a visitor adds another visitor, the currently processed path
is destroyed, all diagnostics is discarded, and it is regenerated again,
until it's no longer modified.
This pattern has a few negative implications:
- This loop does not even guarantee to terminate.
E.g. just imagine two visitors bouncing a diagnostics around.
- Performance-wise, e.g. for sqlite3 all visitors are being re-run at
least 10 times for some bugs.
We have already seen a few reports where it leads to timeouts.
- If we want to add more computationally intense visitors, this will
become worse.
- From architectural standpoint, the current layout requires copying
visitors, which is conceptually wrong, and can be annoying (e.g. no
unique_ptr on visitors allowed).
The proposed change is a much simpler architecture: the outer loop
processes nodes upwards, and whenever the visitor is added it only
processes current nodes and above, thus guaranteeing termination.
Differential Revision: https://reviews.llvm.org/D47856
llvm-svn: 335666
2018-06-27 05:12:08 +08:00
|
|
|
return Out;
|
2008-04-03 12:42:52 +08:00
|
|
|
}
|
2008-07-15 01:40:50 +08:00
|
|
|
|
2012-04-05 02:11:35 +08:00
|
|
|
void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
|
2019-09-07 04:55:29 +08:00
|
|
|
const CheckerBase *Checker, StringRef Name,
|
|
|
|
StringRef Category, StringRef Str,
|
|
|
|
PathDiagnosticLocation Loc,
|
|
|
|
ArrayRef<SourceRange> Ranges,
|
|
|
|
ArrayRef<FixItHint> Fixits) {
|
2019-09-13 03:09:24 +08:00
|
|
|
EmitBasicReport(DeclWithIssue, Checker->getCheckerName(), Name, Category, Str,
|
2019-09-07 04:55:29 +08:00
|
|
|
Loc, Ranges, Fixits);
|
2014-02-12 05:49:21 +08:00
|
|
|
}
|
2018-03-07 08:17:48 +08:00
|
|
|
|
2014-02-12 05:49:21 +08:00
|
|
|
void BugReporter::EmitBasicReport(const Decl *DeclWithIssue,
|
2019-09-13 03:09:24 +08:00
|
|
|
CheckerNameRef CheckName,
|
2014-02-12 05:49:21 +08:00
|
|
|
StringRef name, StringRef category,
|
2011-09-21 05:38:35 +08:00
|
|
|
StringRef str, PathDiagnosticLocation Loc,
|
2019-09-07 04:55:29 +08:00
|
|
|
ArrayRef<SourceRange> Ranges,
|
|
|
|
ArrayRef<FixItHint> Fixits) {
|
2011-02-23 08:16:01 +08:00
|
|
|
// 'BT' is owned by BugReporter.
|
2014-02-12 05:49:21 +08:00
|
|
|
BugType *BT = getBugTypeForName(CheckName, name, category);
|
2019-09-10 04:34:40 +08:00
|
|
|
auto R = std::make_unique<BasicBugReport>(*BT, str, Loc);
|
2012-04-05 02:11:35 +08:00
|
|
|
R->setDeclWithIssue(DeclWithIssue);
|
2019-09-07 04:55:29 +08:00
|
|
|
for (const auto &SR : Ranges)
|
|
|
|
R->addRange(SR);
|
|
|
|
for (const auto &FH : Fixits)
|
|
|
|
R->addFixItHint(FH);
|
2015-06-23 21:15:32 +08:00
|
|
|
emitReport(std::move(R));
|
2009-01-24 04:28:53 +08:00
|
|
|
}
|
2011-02-23 08:16:01 +08:00
|
|
|
|
2019-09-13 03:09:24 +08:00
|
|
|
BugType *BugReporter::getBugTypeForName(CheckerNameRef CheckName,
|
|
|
|
StringRef name, StringRef category) {
|
2012-02-05 10:13:05 +08:00
|
|
|
SmallString<136> fullDesc;
|
2014-02-12 05:49:21 +08:00
|
|
|
llvm::raw_svector_ostream(fullDesc) << CheckName.getName() << ":" << name
|
|
|
|
<< ":" << category;
|
2020-04-29 10:14:58 +08:00
|
|
|
std::unique_ptr<BugType> &BT = StrBugTypes[fullDesc];
|
2014-11-19 11:06:06 +08:00
|
|
|
if (!BT)
|
2020-04-29 10:14:58 +08:00
|
|
|
BT = std::make_unique<BugType>(CheckName, name, category);
|
|
|
|
return BT.get();
|
2011-02-23 08:16:01 +08:00
|
|
|
}
|