1-argument ExecutionEngine::create(Module*) ambiguous with the signature that
used to be ExecutionEngine::create(ModuleProvider*, defaulted_params). Fixed
by removing the 1-argument create(). Fixes PR6221.
llvm-svn: 95236
The SRThreshold value makes perfect sense for checking if an entire aggregate
should be promoted to a scalar integer, but it is not so good for splitting
an aggregate into its separate elements. A struct may contain a large embedded
array along with some scalar fields that would benefit from being split apart
by SROA. Even if the total aggregate size is large, it may still be good to
perform SROA. Thus, the most important piece of this patch is simply moving
the aggregate size comparison vs. SRThreshold so that it guards only the
aggregate promotion.
We have also been checking the number of elements to decide if an aggregate
should be split up. The limit of "SRThreshold/4" seemed rather arbitrary,
and I don't think it's very useful to derive this limit from SRThreshold
anyway. I've collected some data showing that the current default limit of
32 (since SRThreshold defaults to 128) is a reasonable cutoff for struct
types. One thing suggested by the data is that distinguishing between structs
and arrays might be useful. There are (obviously) a lot more large arrays
than large structs (as measured by the number of elements and not the total
size -- a large array inside a struct still counts as a single element given
the way we do SROA right now). Out of 8377 arrays where we successfully
performed SROA while compiling a large set of benchmarks, only 16 of them had
more than 8 elements. And, for those 16 arrays, it's not at all clear that
SROA was actually beneficial. So, to offset the compile time cost of
investigating more large structs for SROA, the patch lowers the limit on array
elements to 8.
This fixes Apple Radar 7563690.
llvm-svn: 95224
prevent a crash on templates when looking for an existing declaration of the
predefined global operators. This fixes PR5918.
Added an easy test case for the overload handling, but testing the crash is
a bit trickier. Created a new test that can use multiple runs with a define to
trigger which test case is used so we can test this type of issue.
llvm-svn: 95220
that is in an anonymous namespace, give that function or variable
internal linkage.
This change models an oddity of the C++ standard, where names declared
in an anonymous namespace have external linkage but, because anonymous
namespace are really "uniquely-named" namespaces, the names cannot be
referenced from other translation units. That means that they have
external linkage for semantic analysis, but the only sensible
implementation for code generation is to give them internal
linkage. We now model this notion via the UniqueExternalLinkage
linkage type. There are several changes here:
- Extended NamedDecl::getLinkage() to produce UniqueExternalLinkage
when the declaration is in an anonymous namespace.
- Added Type::getLinkage() to determine the linkage of a type, which
is defined as the minimum linkage of the types (when we're dealing
with a compound type that is not a struct/class/union).
- Extended NamedDecl::getLinkage() to consider the linkage of the
template arguments and template parameters of function template
specializations and class template specializations.
- Taught code generation to rely on NamedDecl::getLinkage() when
determining the linkage of variables and functions, also
considering the linkage of the types of those variables and
functions (C++ only). Map UniqueExternalLinkage to internal
linkage, taking out the explicit checks for
isInAnonymousNamespace().
This fixes much of PR5792, which, as discovered by Anders Carlsson, is
actually the reason behind the pass-manager assertion that causes the
majority of clang-on-clang regression test failures. With this fix,
Clang-built-Clang+LLVM passes 88% of its regression tests (up from
67%). The specific numbers are:
LLVM:
Expected Passes : 4006
Expected Failures : 32
Unsupported Tests : 40
Unexpected Failures: 736
Clang:
Expected Passes : 1903
Expected Failures : 14
Unexpected Failures: 75
Overall:
Expected Passes : 5909
Expected Failures : 46
Unsupported Tests : 40
Unexpected Failures: 811
Still to do:
- Improve testing
- Check whether we should allow the presence of types with
InternalLinkage (in addition to UniqueExternalLinkage) given
variables/functions internal linkage in C++, as mentioned in
PR5792.
- Determine how expensive the getLinkage() calls are in practice;
consider caching the result in NamedDecl.
- Assess the feasibility of Chris's idea in comment #1 of PR5792.
llvm-svn: 95216
RegionStoreManager::InvalidateRegions() by adjusting the worklist to
iterate over BindingKeys instead of MemRegions. We also only need to
do the actual invalidation work on base regions, and for non-base
regions just blow away their bindings.
llvm-svn: 95200
- Requires backend support, which only exists for i386--darwin currently.
No 'as' required:
--
ddunbar@ozzy:tmp$ cat t.c
int main() { return 42; }
ddunbar@ozzy:tmp$ clang -m32 -integrated-as t.c
ddunbar@ozzy:tmp$ ./a.out; echo $?
42
ddunbar@ozzy:tmp$
--
The random extra whitespace is how you know its working! :)
llvm-svn: 95194
not build a subregion map and instead do a single scan of the store.
This is done by building "region clusters" that represent the
collection of regions that have the same base region. Invalidating
any region in a cluster means that they all should get invalidated.
This change brought out a point that Zhongxing mentioned to me
offline: the flattened memory binding has issues distinguishing
between direct and default bindings. For example, setting the default
value for an entire struct is the same as binding to the first
element. To address this problem, I moved the binding "tag" (Direct
or Default) from BindingVal to BdingKey (and removed BindingVal
entirely). This requires us to do double lookups in some cases; and
there is still much more cleanup that can be done.
This change produced a noticeable speedup when analyzing sqlite3 (a
reduction of 4% in running time).
llvm-svn: 95193
realize that CXXConstructExpr is always implicit, so we should just
return its argument (if there is only one) rather than directly
invoking the constructor.
llvm-svn: 95192
- Don't use GlobalAliases with non-0 GEPs (GNU runtime) - this was unsupported and LLVM will be generating errors if you do it soon. This also simplifies the code generated by the GNU runtime a bit.
- Make GetSelector() return a constant (GNU runtime), not a load of a store of a constant.
- Recognise @selector() expressions as valid static initialisers (as GCC does).
- Add methods to GCObjCRuntime to emit selectors as constants (needed for using @selector() expressions as constants. These need implementing for the Mac runtimes - I couldn't figure out how to do this, they seem to require a load.
- Store an ObjCMethodDecl in an ObjCSelectorExpr so that we can get at the type information for the selector. This is needed for generating typed selectors from @selector() expressions (as GCC does). Ideally, this information should be stored in the Selector, but that would be an invasive change. We should eventually add checks for common uses of @selector() expressions. Possibly adding an attribute that can be applied to method args providing the types of a selector so, for example, you'd do something like this:
- (id)performSelector: __attribute__((selector_types(id, SEL, id)))(SEL)
withObject: (id)object;
Then, any @selector() expressions passed to the method will be check to ensure that it conforms to this signature. We do this at run time on the GNU runtime already, but it would be nice to do it at compile time on all runtimes.
- Made @selector() expressions emit type info if available and the runtime supports it.
Someone more familiar with the Mac runtime needs to implement the GetConstantSelector() function in CGObjCMac. This currently just assert()s.
llvm-svn: 95189
Chris Lattner