Alexey Bataev [Mon, 28 Mar 2016 09:53:43 +0000 (09:53 +0000)]
[OPENMP] Allow runtime insert its own code inside OpenMP regions.
Solution unifies interface of RegionCodeGenTy type to allow insert
runtime-specific code before/after main codegen action defined in
CGStmtOpenMP.cpp file. Runtime should not define its own RegionCodeGenTy
for general OpenMP directives, but must be allowed to insert its own
(required) code to support target specific codegen.
David Majnemer [Mon, 28 Mar 2016 03:19:50 +0000 (03:19 +0000)]
Fix serialization/deserialization for __uuidof
I broke this back in r264529 because I forgot to serialize the UuidAttr
member. Fix this by replacing the UuidAttr with a StringRef which is
properly serialized and deserialized.
Richard Smith [Sun, 27 Mar 2016 20:13:24 +0000 (20:13 +0000)]
[modules] When encoding SourceLocations in bitcode, rotate the 'is macro' flag
bit from the top bit to the bottom bit, so that we don't need 6 VBR6 hunks for
each macro location. Reduces libstdc++ module size by about 1%.
David Majnemer [Sun, 27 Mar 2016 04:46:07 +0000 (04:46 +0000)]
Improve the representation of CXXUuidofExpr
Keep a pointer to the UuidAttr that the CXXUuidofExpr corresponds to.
This makes translating from __uuidof to the underlying constant a lot
more straightforward.
Faisal Vali [Sat, 26 Mar 2016 20:46:45 +0000 (20:46 +0000)]
[NFC] Tweak diagnostic for template template arguments, to include template aliases.
The prior diagnostic (err_template_arg_not_class_template) would state that the template argument to a template template parameter can only be a class template, when it can also be a template alias. The newly renamed diagnostic (err_template_arg_not_valid_template) mentions template aliases.
Bob Wilson [Sat, 26 Mar 2016 18:55:13 +0000 (18:55 +0000)]
Check if a path is already absolute before trying to make it so.
The FileSystem::makeAbsolute function has been calculating the current
working directory unconditionally, even when it is not needed. This calls
down to llvm::sys::fs::current_path, which is relatively expensive
because it stats two directories, regardless of whether those paths are
already in the stat cache. The net effect is that when using the
VFS, every stat during header search turns into three stats. With this
change, we get back to a single stat for absolute directory paths.
Faisal Vali [Sat, 26 Mar 2016 16:11:37 +0000 (16:11 +0000)]
[Cxx1z-constexpr-lambda-P0170R1] Support parsing of constexpr specifier (and its inference) on lambda expressions
Support the constexpr specifier on lambda expressions - and support its inference from the lambda call operator's body.
i.e.
auto L = [] () constexpr { return 5; };
static_assert(L() == 5); // OK
auto Implicit = [] (auto a) { return a; };
static_assert(Implicit(5) == 5);
We do not support evaluation of lambda's within constant expressions just yet.
Implementation Strategy:
- teach ParseLambdaExpressionAfterIntroducer to expect a constexpr specifier and mark the invented function call operator's declarator's decl-specifier with it; Have it emit fixits for multiple decl-specifiers (mutable or constexpr) in this location.
- for cases where constexpr is not explicitly specified, have buildLambdaExpr check whether the invented function call operator satisfies the requirements of a constexpr function, by calling CheckConstexprFunctionDecl/Body.
Much obliged to Richard Smith for his patience and his care, in ensuring the code is clang-worthy.
Richard Smith [Fri, 25 Mar 2016 22:29:27 +0000 (22:29 +0000)]
Don't warn on "use" of undefined inline function that isn't actually an ODR
use. In order for this to fire, the function needed to be a templated function
marked 'constexpr' and declared but not defined. This weird pattern appears in
libstdc++'s alloc_traits.h.
Richard Smith [Fri, 25 Mar 2016 21:49:43 +0000 (21:49 +0000)]
Store list of undefined-but-used objects in a deterministic order to fix
non-deterministic diagnostics (and non-deterministic PCH files). Check these
when building a module rather than serializing it; it's not reasonable for a
module's use to be satisfied by a definition in the user of the module.
Devin Coughlin [Fri, 25 Mar 2016 21:18:22 +0000 (21:18 +0000)]
[analyzer] Add CIFIlter modeling to DeallocChecker.
The -dealloc method in CIFilter is highly unusual in that it will release
instance variables belonging to its *subclasses* if the variable name
starts with "input" or backs a property whose name starts with "input".
Subclasses should not release these ivars in their own -dealloc method --
doing so could result in an over release.
Before this commit, the DeallocChecker would warn about missing releases for
such "input" properties -- which could cause users of the analyzer to add
over releases to silence the warning.
To avoid this, DeallocChecker now treats CIFilter "input-prefixed" ivars
as MustNotReleaseDirectly and so will not require a release. Further, it
will now warn when such an ivar is directly released in -dealloc.
Manman Ren [Fri, 25 Mar 2016 18:43:46 +0000 (18:43 +0000)]
[ObjCXX] Warn undeclared identifiers.
Instantiation dependence were not being handled correctly for OpqaueValueExpr
AST nodes. As a result, if an undeclared identifier was used in a CXXNewExpr
that is assigned to a ObjC property, there would be no error during parsing, and
there would be a crash during code gen. This patch makes sure that an error
will be issued during parsing in this case.
Before the fix, if CXXNewExpr has a typo, its InstantiationDependent will be
set to true, but if it is wrapped in a OpaqueValueExpr, the OpaqueValueExpr will
not be instantiation dependent, causing the TypoExpr not be to resolved. The fix
propagates InstantiationDependent to OpaqueValueExpr from its SourceExpr. It
also propagates the other instantiation bits.
Ben Langmuir [Fri, 25 Mar 2016 17:01:59 +0000 (17:01 +0000)]
[index] Remove redundancy between symbol kind and language
Condense the ObjCKIND and CXXKIND options into just KIND, since the
language was already specified on a per-symbol basis and this
information was redundant. This only changes the internal
representation; naturally the libclang interface remains the same.
Samuel Benzaquen [Fri, 25 Mar 2016 16:29:30 +0000 (16:29 +0000)]
[ASTMatchers] Add own version of VariadicFunction.
Summary:
llvm::VariadicFunction is only being used by ASTMatchers.
Having our own copy here allows us to remove the other one from llvm/ADT.
Also, we can extend the API to meet our needs without modifying the common
implementation.
Adrian Prantl [Fri, 25 Mar 2016 00:20:35 +0000 (00:20 +0000)]
Debug Info: Add a testcase for the bug introduced by r259975.
In r259975 we rauw'ed the scope of enum declarations without taking into
account that DIBuilder strips out scope references that point to the
DICompileUnit to facilitate type uniquing.
This testcase guards against making the same mistake again.
<rdar://problem/25078246>
Richard Smith [Thu, 24 Mar 2016 23:41:14 +0000 (23:41 +0000)]
[modules] Store offset to LOCAL_REDECLARATIONS record relative to the current
record rather than relative to the start of the bitcode file. Saves a couple of
bytes per LOCAL_REDECLARATIONS record (also makes diffs of llvm-bcanalyzer
output more useful when tracking down nondeterminism...).
Reid Kleckner [Thu, 24 Mar 2016 20:38:43 +0000 (20:38 +0000)]
Revert "Recommitted r263425 "Supporting all entities declared in lexical scope in LLVM debug info." After fixing PR26942 (the fix is included in this commit)."
This reverts commit r264281.
This change breaks building Chromium for iOS. We'll upload a reproducer
to the PR soon.
Richard Smith [Thu, 24 Mar 2016 19:12:22 +0000 (19:12 +0000)]
Change ADL to produce lookup results in a deterministic order. This fixes some
rare issues with nondeterministic diagnostic order, and some very common issues
with nondeterministic module builds.
Eric Liu [Thu, 24 Mar 2016 13:22:37 +0000 (13:22 +0000)]
Revert "Added support for different VFSs in format::getStyle."
This reverts commit r264253. It is breaking the buildbot http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-windows10pro-fast/builds/2203
Eric Liu [Thu, 24 Mar 2016 11:25:28 +0000 (11:25 +0000)]
Revert "Added support for different VFSs in format::getStyle."
This reverts commit r264253. It is breaking the buildbot http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-windows10pro-fast/builds/2203
Eric Liu [Thu, 24 Mar 2016 10:50:17 +0000 (10:50 +0000)]
Added support for different VFSs in format::getStyle.
Summary:
Previously, format::getStyle assumes that the given file resides in
the real file system, which prevents the use of virtual file system in testing etc.
This patch adds a parameter in format::getStyle interface so that users can specify
the right file system. By default, the file system is the real file system.
Eric Christopher [Thu, 24 Mar 2016 00:34:02 +0000 (00:34 +0000)]
Modules builds are necessarily compile actions, but they don't
necessarily produce object files. Turn off split dwarf if we're not
producing a file that the driver believes is an object file.
Nico Weber [Wed, 23 Mar 2016 22:57:55 +0000 (22:57 +0000)]
clang-cl: With -fmsc-version=1900, use MSVS2015 diag formatting.
Remove tests that have neither a triple nor an explicit -fmsc-version flag,
since in the absence of an -fmsc-version flag, the implicit value of the flag
is 17 (MSVC2013) with MSVC triples but 0 (not set) for other triples, and
the default triple is platform dependent.
Justin Lebar [Wed, 23 Mar 2016 22:42:30 +0000 (22:42 +0000)]
[CUDA] Merge most of CodeGenCUDA/function-overload.cu into SemaCUDA/function-overload.cu.
Summary:
Previously we were using the codegen test to ensure that we choose the
right overload. But we can do this within sema, with a bit of
cleverness.
I left the constructor/destructor checks in CodeGen, because these
overloads (particularly on the destructors) are hard to check in Sema.
Summary:
Principally, don't hardcode the line numbers of various notes. This
lets us make changes to the test without recomputing linenos everywhere.
Instead, just tell -verify that we may get 0 or more notes pointing to
the relevant function definitions. Checking that we get exactly the
right note isn't so important (and anyway is checked elsewhere).
Justin Lebar [Wed, 23 Mar 2016 22:42:27 +0000 (22:42 +0000)]
[CUDA] Don't define __NVCC__.
Summary:
We decided this makes life too difficult for code authors. For example,
people may want to detect NVCC and disable variadic templates, which
NVCC does not support, but which we do.
Since people are going to have to change compiler flags *anyway* in
order to compile with clang, if they really want the old behavior, they
can pass -D__NVCC__.
Tested with tensorflow and thrust, no apparent problems.
Richard Smith [Wed, 23 Mar 2016 20:39:06 +0000 (20:39 +0000)]
Make SemaAccess smarter about determining when a dependent class might
instantiate to match a friend class declaration. It's still pretty dumb,
though.
Richard Smith [Wed, 23 Mar 2016 20:07:07 +0000 (20:07 +0000)]
Make sure to perform dependent access checks when instantiating a
lambda-expression. We don't actually instantiate the closure type / operator()
in the template in order to produce the closure type / operator() in the
instantiation, so this isn't caught by the normal path.
Nico Weber [Wed, 23 Mar 2016 18:46:57 +0000 (18:46 +0000)]
clang-cl: Fix remaining bugs in interaction of /Yc and /FI /showIncludes.
Instead of putting the /Yc header into ExtraDeps, give DependencyOutputOptions
a dedicated field for /Yc mode, and let HeaderIncludesCallback hang on to the
full DependencyOutputOptions object, not just ExtraDeps.
Reverts parts of r263352 that are now no longer needed.
Nico Weber [Wed, 23 Mar 2016 18:17:02 +0000 (18:17 +0000)]
clang-cl: Add more tests for the interaction of /FI and /Yc /Yu.
Most things even work; see the included FIXMEs for things that need polishing.
Also don't warn about unused flags for the `/Yuh2.h /FIh1.h /FIh2.h`. The
common case is that the pch was built with `/Ych2.h /FIh1.h /FIh2.h`, so h1.h
is in the PCH, and we shouldn't warn about /FIh1.h not having an effect.
(If we wanted to get fancy, we could store the list of -include flags in the
pch and then check that it matches later on.)
Nico Weber [Wed, 23 Mar 2016 18:00:22 +0000 (18:00 +0000)]
clang-cl: Include /FI headers in /showIncludes output.
-H in gcc mode doesn't print -include headers, but they are included in
depfiles written by MMD and friends. Since /showIncludes is what's used instead
of depfiles, printing /FI there seems important (and matches cl.exe).
Instead of giving HeaderIncludeGen more options, just switch on ShowAllHeaders
in clang-cl mode and let clang::InitializePreprocessor() not put -include flags
in the <command line> block. This changes the behavior of -E slightly, and it
removes the <command line> flag from the output triggered by setting the
obscure CC_PRINT_HEADERS=1 env var to true while running clang. Both of these
seem ok to change.
Manman Ren [Wed, 23 Mar 2016 16:28:28 +0000 (16:28 +0000)]
ObjC: Handle boolean fixed type for enum.
Before this commit, we assert failure in ImplicitCastExpr
"unheralded conversion to bool". This commit fixes the assertion by using
the correct cast type when the fixed type is boolean.
This commit also fixes the behavior for Microsoft mode as well, since
Obj-C and Microsoft mode share the same code path.
void *fp = f; // OK. This is C and the target is void*.
void (*fp2)(void) = f; // Error. This is C, but the target isn't void*.
```
This patch makes the assignment of `fp2` select the `f(int)` overload,
rather than emitting an error (N.B. you'll still get a warning about the
`fp2` assignment if you use -Wincompatible-pointer-types).
Justin Bogner [Tue, 22 Mar 2016 17:50:05 +0000 (17:50 +0000)]
StaticAnalyzer: Avoid an unintentional copy
The range here isn't over references, so using `auto &` here incites a
copy. Switching to `auto *` would do, but we might as well list an
explicit type for clarity.
Summary: A checker (will be uploaded after this patch) needs to check implicit casts. Existing generic matcher "has" ignores implicit casts and parenthesized expressions and no specific matcher for matching return value expression preexisted. The patch adds such a matcher (hasReturnValue).
Akira Hatanaka [Tue, 22 Mar 2016 05:00:21 +0000 (05:00 +0000)]
[Objective-c] Do not set IsExact to true when the receiver is a class.
IsExact shouldn't be set to true in WeakObjectProfileTy::getBaseInfo
when the receiver is a class because having a class as the receiver
doesn't guarantee that the Base is exact.
Mike Spertus [Tue, 22 Mar 2016 02:29:22 +0000 (02:29 +0000)]
Visual Studio Native Visualizations for constructors and methods
With this change, the class
struct A {
A(int _i);
~A();
int foo(double d);
double bar(A *a) { return 1.3; }
};
appears in the VS2015 Locals Window as
D 0x02dbb378 struct A
|- DeclKind CXXRecord
|- Members
|- [0] implicit struct A
|- [1] Constructor {A(int _i)}
|- [2] Destructor {~A()}
|- [3] Method {int foo(double d)}
|- [4] Method {double bar(struct A *)}
|- [Raw View] /* Other stuff */
Note that these changes only benefit VS2015 as
VS2013 does not have views and only displays the
struct name "A", but the change does no apparent
harm in VS2013, so is still a win.
Justin Lebar [Tue, 22 Mar 2016 00:09:28 +0000 (00:09 +0000)]
[CUDA] Implement atomicInc and atomicDec builtins
These functions cannot be implemented as atomicrmw or cmpxchg
instructions, so they are implemented as a call to the NVVM intrinsics
@llvm.nvvm.atomic.load.inc.32.p0i32 and
@llvm.nvvm.atomic.load.dec.32.p0i32.
Chris Bieneman [Mon, 21 Mar 2016 22:37:14 +0000 (22:37 +0000)]
[Perf-training] Adding support for tests to skip the clang driver
This patch adds a new set of substitutions to the lit run lines for order files and PGO generation which run the clang driver to get the cc1 command, then execute the cc1 command directly. This allows the scripts to bypass profiling the clang driver over and over again.
The approach in this patch was discussed via IRC with Sean Silvas.
Special thanks to Daniel Dunbar whose out-of-tree code I liberally plagiarized.
Richard Smith [Mon, 21 Mar 2016 21:16:01 +0000 (21:16 +0000)]
[modules] Renumber DECL_UPDATES from 30 to 50, so it no longer collides with
TYPE_TEMPLATE_SPECIALIZATION. This was fine in practice because both record
kinds are only ever found by offset, but made the llvm-bcanalyzer -dump output
very confusing.
Pete Cooper [Mon, 21 Mar 2016 20:50:03 +0000 (20:50 +0000)]
Revert "Convert some ObjC msgSends to runtime calls."
This reverts commit r263607.
This change caused more objc_retain/objc_release calls in the IR but those
are then incorrectly optimized by the ARC optimizer. Work is going to have
to be done to ensure the ARC optimizer doesn't optimize user written RR, but
that should land before this change.
This change will also need to be updated to take account for any changes required
to ensure that user written calls to RR are distinct from those inserted by ARC.
projects / clang / log