Add member template 'Type::getAs<T>', which converts a Type* to a respective T*.
This method is intended to eventually replace the individual
Type::getAsXXXType<> methods.
The motivation behind this change is twofold:
1) Reduce redundant implementations of Type::getAsXXXType() methods. Most of
them are basically copy-and-paste.
2) By centralizing the implementation of the getAs<Type> logic we can more
smoothly move over to Doug Gregor's proposed canonical type smart pointer
scheme.
Along with this patch:
a) Removed 'Type::getAsPointerType()'; now clients use getAs<PointerType>.
b) Removed 'Type::getAsBlockPointerTypE()'; now clients use getAs<BlockPointerType>.
Chris Lattner [Thu, 16 Jul 2009 16:48:25 +0000 (16:48 +0000)]
use CreateRuntimeVariable to get __CFConstantStringClassReference
instead of doing it manually (which might end up getting auto-renamed),
fixing a FIXME. rdar://7065446
Move RegionStoreManager over to using new
ValueManager::makeArrayIndex()/convertArrayIndex() methods. This
handles yet another crash case when reasoning about array indices of
different bitwidth and signedness.
(1) Moved the SValuator object from GRExprEngine to ValueManager. This
allows ValueManager to use the SValuator when creating SVals.
(2) Added ValueManager::makeArrayIndex() and
ValueManager::convertToArrayIndex(), two SVal creation methods
that will help RegionStoreManager always have a consistent set of
SVals with the same integer size and type when reasoning about
array indices.
Handle some more fallout with the conversion of using PointerType for
Objective-C pointers to using ObjCObjectPointerType.
Now the checking for 'attribute ((nonnull))' in Sema doesn't emit an error when
trying to apply that attribute to a parameter that is an Objective-C pointer
(this is a regression).
To prevent this regression from occuring in the future, the 'nonnull.c' test was
moved to test/SemaObjC and renamed 'nonnull.m'. I also enhanced the tests to
show that function calls involved a NULL Objective-C pointer constant does not
trigger a warning. This is consistent with GCC, but should likely be fixed.
Steve Naroff [Wed, 15 Jul 2009 18:40:39 +0000 (18:40 +0000)]
Implement the ObjC pseudo built-in types as clang "BuiltinType's". I say pseudo built-in types, since Sema still injects a typedef for recognition (i.e. they aren't truly built-ins from a parser perspective).
This removes the static data/methods on ObjCObjectPointerType while preserving the nice API (no need to fiddle with ASTContext:-).
This patch also adds Type::isObjCBuiltinType().
This should be the last fairly large patch related to recrafting the ObjC type system. The follow-on patches should be fairly small.
Dan Gohman [Wed, 15 Jul 2009 17:32:18 +0000 (17:32 +0000)]
Update for raw_fd_ostream API changes. raw_fd_ostream now has a
Force flag to control whether the case of opening an existing
file is considered an error.
More test cases revealed that the logic in StoreManager::InvalidateRegion() needs more finesse when handling the invalidation of pointers. Pointers that were invalidated as integers could later cause problems for clients using them as pointers. It is easier for us to model a symbolic value as a pointer rather than modeling a non-symbolic value as a pointer.
This patch causes:
- StoreManager::InvalidateRegion() to not used the casted type of a region if
it would cause a pointer type to be invalidated as a non-pointer type.
- Pushes RegionStore::RetrieveElement() further by handling retrievals from
symbolic arrays that have been invalidated. This uses the new SymbolDerived
construct that was recently introduced.
The result is that the failing test in misc-ps-region-store-x86_64.m now passes.
Both misc-ps-region-store-x86_64.m and misc-ps-region-store-i386.m contain a
test case that motivated this change.
Introduced the notion of a "derived symbol" using the class SymbolDerived.
SymbolDerived allows us to model symbolic values that are related to other
symbols via a region hierarchy. For example, SymbolDerived can be used to model
individual values of a symbolic array.
Split out 'test2' into an i386 and x86_64 file, illustrating how the
test behavior differs between architectures. When this is no longer
the case, these tests will be merged.
Eli Friedman [Tue, 14 Jul 2009 21:58:17 +0000 (21:58 +0000)]
Remove -ftraditional option, which gcc doesn't actually support. Make
using -traditional and -traditional-cpp with clang an error because
it's unsupported in clang and causes a significant change in the
semantics of the language.
Enhance RegionStoreManager to handle 'Retrieve's from SymbolicRegions. We do this by silently wrapping the region with an ElementRegion. This fixes the failures in misc-ps-region-store.m.
Steve Naroff [Tue, 14 Jul 2009 14:58:18 +0000 (14:58 +0000)]
Add a "TypeSpecStartLoc" to FieldDecl. Patch contributed by Enea Zaffanella.
Note: One day, it might be useful to consider adding this info to DeclGroup (as the comments in FunctionDecl/VarDecl suggest). For now, I think this works fine. I considered moving this to ValueDecl (a common ancestor of FunctionDecl/VarDecl/FieldDecl), however this would add overhead to EnumConstantDecl (which would burn memory and isn't necessary).
Introduce redecl_iterator, used for iterating over the redeclarations of a FunctionDecl or VarDecl.
It iterates over all the redeclarations, regardless of the starting point. For example:
1) int f();
2) int f();
3) int f();
if you have the (2) FunctionDecl and call redecls_begin/redecls_end to iterate, you'll get this sequence:
(2)
(1)
(3)
The motivation to introduce this was that, previously, if (3) was a function definition,
and you called getBody() at (2), it would not return it, since getBody() iterated over the previous declarations only,
so it would only check (2) and (1).
- Declaration context of ParmVarDecls (that we got from the Declarator) was not their containing function.
- C++ out-of-line method definitions didn't get an access specifier.
Both were exposed by a crash when emitting a C++ method to a PCH file (assert at Decl::CheckAccessDeclContext()).
Daniel Dunbar [Tue, 14 Jul 2009 01:20:56 +0000 (01:20 +0000)]
Update debug info generation for ObjCObjectPointer changes.
- Previously this would crash on recursive types, and it was also incorrectly
stripping off a level of indirection.
- I'm not 100% convinced this is all correct, but it should be a monotonic
improvment.
Instead of recovering from a wrong invalidation, this patch aims to
invalidate the region correctly. It uses the cast-to type to invalidate
the region when available. To avoid invalid cast-to type like 'void*' or 'id',
region store now only records non-generic casts of regions.
Add basic checking for passing NULL to CFRetain/CFRelease, since those functions
are not explicitly marked as not accepting NULL pointers. This check illustrates
how we need more refactoring in the custom-check logic.
Enhance SimpleSValuator::EvalBinOpNN to recognize the trivial case
where we are comparing a symbolic value against itself, regardless of
the nature of that symbolic value.
This enhancement identified a case where RegionStoreManager is not
correctly symbolicating the values of the pointees of parameters. The
failing test is now in 'test/Analysis/misc-ps-region-store.m', with
that test file now (temporarily) marked XFAIL.
projects / clang / log