Douglas Gregor [Thu, 23 Jul 2009 16:36:45 +0000 (16:36 +0000)]
Clean up the ActOnTag action, so that there is only a single entry
point that covers templates and non-templates. This should eliminate
the flood of warnings I introduced yesterday.
Removed the ActOnClassTemplate action, which is no longer used.
Add template specializations to view the call graph in dot format.
- change the DenseMap used in callgraph to std::map, since DenseMap cannot
be used with mapped_iterator and friends.
Daniel Dunbar [Thu, 23 Jul 2009 05:01:54 +0000 (05:01 +0000)]
Use llvm::BitVector instead of managing memory by hand.
- As it happens, this also fixes a use-of-uninitialized memory that was causing
non-deterministic test failures.
Anders Carlsson [Thu, 23 Jul 2009 04:50:01 +0000 (04:50 +0000)]
Use arrays as union padding. Also, since the resulting struct will always contain a single element and either a single i8 element or an array of i8s, there's no reason to use a packed struct.
Anders Carlsson [Thu, 23 Jul 2009 03:17:50 +0000 (03:17 +0000)]
Check in CGRecordLayoutBuilder which is a reimplementation of the record layout code. (Yay, no more packed structs unless absolutely necessary). We currently don't use the layouts being built but that will change when the new code is mature enough :)
Steve Naroff [Thu, 23 Jul 2009 01:01:38 +0000 (01:01 +0000)]
Remove a bunch of FIXME's related to ObjC type checking.
- Move Sema::ObjCQualifiedIdTypesAreCompatible(), Sema::QualifiedIdConformsQualifiedId(), and a couple helper functions to ASTContext.
- Change ASTContext::canAssignObjCInterfaces() to use ASTContext:: ObjCQualifiedIdTypesAreCompatible().
- Tweak several test cases to accommodate the new/improved type checking.
Mike Stump [Wed, 22 Jul 2009 23:56:57 +0000 (23:56 +0000)]
Add warning for falling off the end of a function that should return a
value. This is on by default, and controlled by -Wreturn-type (-Wmost
-Wall). I believe there should be very few false positives, though
the most interesting case would be:
int() { bar(); }
when bar does:
bar() { while (1) ; }
Here, we assume functions return, unless they are marked with the
noreturn attribute. I can envision a fixit note for functions that
never return normally that don't have a noreturn attribute to add a
noreturn attribute.
If anyone spots other false positives, let me know!
Implementing this feature required some extensions to ActOnTag, which
now takes a set of template parameter lists, and is the precursor to
removing the ActOnClassTemplate function from the parser Action
interface. The reason for this approach is simple: the parser cannot
tell the difference between a class template definition and the
definition of a member of a class template; both have template
parameter lists, and semantic analysis determines what that template
parameter list means.
There is still some cleanup to do with ActOnTag and
ActOnClassTemplate. This commit provides the basic functionality we
need, however.
Refactor 'PostStmt' and 'PreStmt' to subclass a common parent 'StmtPoint'.
Educate GRExprEngine::VisitGraph() about 'PreStmt'.
Mark the constructor of 'PostStmt' to be explicit, preventing implicit
conversions and the selection of the wrong 'generateNode' method in
GRStmtNodeBuilder.
Constify a bunch of arguments, which falls out of the changes to ProgramPoint.
Migrate the path-sensitive checking of 'nonnull' arguments over to the new
'Checker' interface. An updated test case illustrates that after calling a
function with the 'nonnull' attribute we now register the fact that the passed
pointer must be non-null. This retention of information was not possible with
the previously used GRSimpleAPICheck interface.
Add support for registering 'Checker' objects with GRExprEngine.
Add a 'previsit' stage (that dispatches to registered Checkers) when evaluating the effects of CallExprs.
Add new 'Checker' and 'CheckerVisitor' classes, which represent a more powerful
interface to plug in domain-specific checker logic than the current
GRSimpleAPICheck interface. The new 'Checker' interface can actually generate
new nodes, allowing it to modify the state and refine the analysis (which
GRSimpleAPIChecks could not).
Douglas Gregor [Wed, 22 Jul 2009 21:30:48 +0000 (21:30 +0000)]
Fix the problems with template argument deduction and array types for
real. It turns out that we need to actually move all of the qualifiers
up to the array type itself, then recanonicalize the deduced template
argument type.
Douglas Gregor [Wed, 22 Jul 2009 17:18:37 +0000 (17:18 +0000)]
Implement parsing and semantic analysis for out-of-line definitions of static
data members of class templates. We don't instantiate the definitions yet,
however.
Restructure ProgramPoint to have the 'Kind' value be its own instance
variable. This gives us much more flexibility with defining more
ProgramPoints, which is the direction we are heading. The removal of
various bit-mangling of pointers also cleans up the logic.
Temporarily disable most use of region casts in RegionStoreManager,
instead preferring to use the a region's actual type when creating
symbols and using the QualType passed to Retrieve for implicit
casting. This preprocessor logic is temporary; eventually we will
either remove region casts or keep them.
Temporarily toggle (via preprocessor directives) that SymbolicRegions
with heap storage are not undefined, but instead should be
symbolicated. If we want to model that a SymbolicRegion is
uninitialized, we can explicitly model that by binding UndefinedVal to
that region. It turns out that we want to treat most heap objects as
being defined, not the other way around.
Fix a crasher in StoreManager::InvalidateRegion() caused by using the
'cast type' of a region to invalidate its binding. This only occurs
when using RegionStoreManager, as it records the cast type. I'm
currently considering removing the notion of a cast type (see
comments in code).
Mon P Wang [Wed, 22 Jul 2009 03:08:17 +0000 (03:08 +0000)]
Preserve address space information through member accesses, e.g.,
__attribute__((address_space(1))) struct {int arr[ 3 ]; } *p1;
... = p1->arr[2]; // load from address space 1
Douglas Gregor [Wed, 22 Jul 2009 00:28:09 +0000 (00:28 +0000)]
Complain if we're entering the context of a dependent nested-name-specifier but
cannot match that nested-name-specifier to a class template or class template
partial specialization.
Fix long-lurking bug in ObjCSummaryCache revealed by Torok's recent memory
poisoning changes to DenseMap. We were using an iterator after it had been
invalidated by an insertion into the DenseMap.
Fix PR 4594 by refactoring almost all casting logic from GRExprEngine::VisitCast
to SValuator::EvalCast. In the process, the StoreManagers now use this new cast
machinery, and the hack in GRExprEngine::EvalBind to handle implicit casts
involving OSAtomicCompareAndSwap and friends has been removed (and replaced with
logic closer to the logic specific to those functions).
Eli Friedman [Tue, 21 Jul 2009 19:28:10 +0000 (19:28 +0000)]
Misc fixes for -Wreorder:
1. Make it work correctly with anonymous unions.
2. Don't compute it if the warning isn't enabled.
3. Optimize the algorithm slightly to make it linear time in the
case where we don't produce any warnings.
projects / clang / log