Roman Lebedev [Thu, 9 May 2019 10:48:13 +0000 (10:48 +0000)]
Revert "[OPENMP]Fix PR41768: check DSA for globals with `default(none)` clauses."
This has introduced (exposed?) a crash in clang sema,
that does not happen without this patch.
I'll followup in the original bugreport and commit with reproducer.
Roman Lebedev [Thu, 9 May 2019 10:47:45 +0000 (10:47 +0000)]
Revert "[OPENMP]Fix PR41767: diagnose DSA for variables in clauses with default(none)."
This implementation isn't sound as per the standard.
It erroneously diagnoses e.g. the following case:
```
$ cat test.cpp
void f(int n) {
#pragma omp parallel default(none) if(n)
;
}
```
```
$ ./bin/clang -fopenmp test.cpp
test.cpp:2:40: error: variable 'n' must have explicitly specified data sharing attributes
#pragma omp parallel default(none) if(n)
^
test.cpp:2:31: note: explicit data sharing attribute requested here
#pragma omp parallel default(none) if(n)
^
1 error generated.
```
As per OpenMP Application Programming Interface Version 5.0 November 2018:
* 2.19.4.1default Clause
The default clause explicitly determines the data-sharing attributes of
variables that are referenced *in a parallel, teams, or task generating
construct and would otherwise be implicitly determined
(see Section 2.19.1.1 on page 270).
* 2.6.1 Determining the Number of Threads for a parallel Region
Using a variable in an if or num_threads clause expression of a parallel
construct causes an implicit reference to the variable in all enclosing
constructs. The if clause expression and the num_threads clause expression
are evaluated in the context outside of the parallel construct,
Summary:
When using `clang -mcpu=CPUNAME+FEATURELIST`,
the implied features defined by CPUNAME are
not obtained, as the entire string is passed.
This fixes that by spiting the cpuname
string in the first `+`, if any.
For example, when using
```clang -### --target=arm-arm-none-eabi -march=armv7-a -mcpu=cortex-a8+nocrc```
the intrinsic
```"target-feature" "+dsp"```
implied by `cortex-a8` is missing.
Petr Hosek [Thu, 9 May 2019 06:09:35 +0000 (06:09 +0000)]
[NewPM] Setup Passes for KASan and KMSan
While ASan and MSan passes were already ported to new PM, the kernel
variants weren't setup in the pipeline which makes the KASan and KMSan
tests in Clang fail.
Richard Smith [Thu, 9 May 2019 03:31:27 +0000 (03:31 +0000)]
[c++20] Implement P0846R0: allow (ADL-only) calls to template-ids whose
template name is not visible to unqualified lookup.
In order to support this without a severe degradation in our ability to
diagnose typos in template names, this change significantly restructures
the way we handle template-id-shaped syntax for which lookup of the
template name finds nothing.
Instead of eagerly diagnosing an undeclared template name, we now form a
placeholder template-name representing a name that is known to not find
any templates. When the parser sees such a name, it attempts to
disambiguate whether we have a less-than comparison or a template-id.
Any diagnostics or typo-correction for the name are delayed until its
point of use.
The upshot should be a small improvement of our diagostic quality
overall: we now take more syntactic context into account when trying to
resolve an undeclared identifier on the left hand side of a '<'. In
fact, this works well enough that the backwards-compatible portion (for
an undeclared identifier rather than a lookup that finds functions but
no function templates) is enabled in all language modes.
Mike Rice [Wed, 8 May 2019 17:15:21 +0000 (17:15 +0000)]
Include corecrt.h in stddef.h and vcruntime.h in stdarg.h to improve MS
compatibility. This allows some applications developed with MSVC to
compile with clang without any extra changes.
Summary:
In this patch we propose a temporary solution to resolving math functions for the NVPTX toolchain, temporary until OpenMP variant is supported by Clang.
We intercept the inclusion of math.h and cmath headers and if we are in the OpenMP-NVPTX case, we re-use CUDA's math function resolution mechanism.
[Sema][OpenCL] Make address space conversions a bit stricter.
The semantics for converting nested pointers between address
spaces are not very well defined. Some conversions which do not
really carry any meaning only produce warnings, and in some cases
warnings hide invalid conversions, such as 'global int*' to
'local float*'!
This patch changes the logic in checkPointerTypesForAssignment
and checkAddressSpaceCast to fail properly on implicit conversions
that should definitely not be permitted. We also dig deeper into the
pointer types and warn on explicit conversions where the address
space in a nested pointer changes, regardless of whether the address
space is compatible with the corresponding pointer nesting level
on the destination type.
Richard Smith [Wed, 8 May 2019 01:36:36 +0000 (01:36 +0000)]
Split ActOnCallExpr into an ActOnCallExpr to be called by the parser,
and a BuildCallExpr to be called internally within Sema to build /
rebuild calls.
Reid Kleckner [Tue, 7 May 2019 23:06:21 +0000 (23:06 +0000)]
[COFF] Use COFF stubs for extern_weak functions
Summary:
A COFF stub indirects the reference to a symbol through memory. A
.refptr.$sym global variable pointer is created to refer to $sym.
Typically mingw uses these for external global variable declarations,
but we can use them for weak function declarations as well.
Updates the dso_local classification to add a special case for
extern_weak symbols on COFF in both clang and LLVM.
Artem Dergachev [Tue, 7 May 2019 22:33:13 +0000 (22:33 +0000)]
[analyzer] Fix a crash when doing RVO from within blocks.
When looking for the location context of the call site, unwrap block invocation
contexts because they are attached to the current AnalysisDeclContext
while what we need is the previous AnalysisDeclContext.
David Blaikie [Tue, 7 May 2019 21:38:51 +0000 (21:38 +0000)]
-frewrite-imports: Add support for wildcard rules in umbrella modules with
This trips over a few other limitations, but in the interests of incremental development I'm starting here & I'll look at the issues with -verify and filesystem checks (the fact that the behavior depends on the existence of a 'foo' directory even though it shouldn't need it), etc.
Revert "[OpenMP][Clang] Support for target math functions"
This commit appears to be breaking stage-2 builds on GreenDragon. The
OpenMP wrappers for cmath and math.h are copied into the root of the
resource directory and cause a cyclic dependency in module 'Darwin':
Darwin -> std -> Darwin. This blows up when CMake is testing for modules
support and breaks all stage 2 module builds, including the ThinLTO bot
and all LLDB bots.
CMake Error at cmake/modules/HandleLLVMOptions.cmake:497 (message):
LLVM_ENABLE_MODULES is not supported by this compiler
Richard Smith [Tue, 7 May 2019 07:36:07 +0000 (07:36 +0000)]
Improve function / variable disambiguation.
Keep looking for decl-specifiers after an unknown identifier. Don't
issue diagnostics about an error type specifier conflicting with later
type specifiers.
Petr Hosek [Mon, 6 May 2019 23:24:17 +0000 (23:24 +0000)]
[Clang][NewPM] Don't bail out if the target machine is empty
This matches the behavior of the old pass manager. There are some
targets that don't have target machine at all (e.g. le32, spir) which
whose tests would never run with new pass manager. Similarly, we would
need to disable tests for targets that are disabled.
Martin Storsjo [Mon, 6 May 2019 21:19:07 +0000 (21:19 +0000)]
[AArch64] Add __builtin_sponentry, for calling setjmp in MinGW
In MinGW, setjmp isn't expanded as a builtin in the compiler (like it
is for MSVC), but manually hooked up as calls to the right underlying
functions in headers. Using the actual CRT's real setjmp/longjmp
functions requires this intrinsic. (Currently this is worked around by
using MinGW specific reimplementations of setjmp/longjmp on aarch64.)
Alexey Bataev [Mon, 6 May 2019 20:07:20 +0000 (20:07 +0000)]
[OPENMP]Fix PR41767: diagnose DSA for variables in clauses with
default(none).
If the combined directive has default(none) clause and has clauses for
inner directive that reference some variables, for which data-sharing
attributes are not specified, the error messages should be emitted for
such variables.
Summary:
In this patch we propose a temporary solution to resolving math functions for the NVPTX toolchain, temporary until OpenMP variant is supported by Clang.
We intercept the inclusion of math.h and cmath headers and if we are in the OpenMP-NVPTX case, we re-use CUDA's math function resolution mechanism.
Alexey Bataev [Mon, 6 May 2019 17:49:22 +0000 (17:49 +0000)]
[OPENMP]Fix PR41768: check DSA for globals with `default(none)` clauses.
If the `default(none)` was specified for the construct, we might miss
diagnostic for the globals without explicitly specified data-sharing
attributes. Patch fixes this problem.
Ilya Biryukov [Mon, 6 May 2019 13:18:00 +0000 (13:18 +0000)]
[CodeComplete] Add a trailing semicolons to some pattern completions
Summary:
Where semicolon is required in any case. Here's a list of completions
that now have a semicolon:
- namespace <name> = <target>;
- using namespace <name>;
- using <qualifier>::<name>;
- continue;
- break;
- goto <label>;
- return;
- return <expression>;
Hans Wennborg [Mon, 6 May 2019 09:51:10 +0000 (09:51 +0000)]
Revert r359949 "[clang] adding explicit(bool) from c++2a"
This caused Clang to start erroring on the following:
struct S {
template <typename = int> explicit S();
};
struct T : S {};
struct U : T {
U();
};
U::U() {}
$ clang -c /tmp/x.cc
/tmp/x.cc:10:4: error: call to implicitly-deleted default constructor of 'T'
U::U() {}
^
/tmp/x.cc:5:12: note: default constructor of 'T' is implicitly deleted
because base class 'S' has no default constructor
struct T : S {};
^
1 error generated.
See discussion on the cfe-commits email thread.
This also reverts the follow-ups r359966 and r359968.
> this patch adds support for the explicit bool specifier.
>
> Changes:
> - The parsing for the explicit(bool) specifier was added in ParseDecl.cpp.
> - The storage of the explicit specifier was changed. the explicit specifier was stored as a boolean value in the FunctionDeclBitfields and in the DeclSpec class. now it is stored as a PointerIntPair<Expr*, 2> with a flag and a potential expression in CXXConstructorDecl, CXXDeductionGuideDecl, CXXConversionDecl and in the DeclSpec class.
> - Following the AST change, Serialization, ASTMatchers, ASTComparator and ASTPrinter were adapted.
> - Template instantiation was adapted to instantiate the potential expressions of the explicit(bool) specifier When instantiating their associated declaration.
> - The Add*Candidate functions were adapted, they now take a Boolean indicating if the context allowing explicit constructor or conversion function and this boolean is used to remove invalid overloads that required template instantiation to be detected.
> - Test for Semantic and Serialization were added.
>
> This patch is not yet complete. I still need to check that interaction with CTAD and deduction guides is correct. and add more tests for AST operations. But I wanted first feedback.
> Perhaps this patch should be spited in smaller patches, but making each patch testable as a standalone may be tricky.
>
> Patch by Tyker
>
> Differential Revision: https://reviews.llvm.org/D60934
Luo, Yuanke [Mon, 6 May 2019 08:25:11 +0000 (08:25 +0000)]
Enable intrinsics of AVX512_BF16, which are supported for BFLOAT16 in Cooper Lake
Summary:
1. Enable infrastructure of AVX512_BF16, which is supported for BFLOAT16 in Cooper Lake;
2. Enable intrinsics for VCVTNE2PS2BF16, VCVTNEPS2BF16 and DPBF16PS instructions, which are Vector Neural Network Instructions supporting BFLOAT16 inputs and conversion instructions from IEEE single precision.
For more details about BF16 intrinsic, please refer to the latest ISE document: https://software.intel.com/en-us/download/intel-architecture-instruction-set-extensions-programming-reference
Richard Smith [Mon, 6 May 2019 03:47:15 +0000 (03:47 +0000)]
[c++20] Implement P1009R2: allow omitting the array bound in an array
new expression.
This was voted into C++20 as a defect report resolution, so we
retroactively apply it to all prior language modes (though it can never
actually be used before C++11 mode).
The 'CodeGenObjC/illegal-UTF8.m' get failed with Clang built with 32-bit targets only (as example ARM-only) with the following error:
error: unable to create target: 'No available targets are compatible with triple "< ... any 64-bit target triple ... >"'
I didn't find any 64-bit dependencies for the test and I think removing '-m64' option should fix the problem and allow this test for any target specified by LLVM_DEFAULT_TARGET_TRIPLE.
Kristof Umann [Sun, 5 May 2019 19:42:33 +0000 (19:42 +0000)]
[analyzer][UninitializedObjectChecker] PR41741: Regard all scalar types as primitive.
https://bugs.llvm.org/show_bug.cgi?id=41741
Pretty much the same as D61246 and D61106, this time for __complex__ types. Upon
further investigation, I realized that we should regard all types
Type::isScalarType returns true for as primitive, so I merged
isMemberPointerType(), isBlockPointerType() and isAnyComplexType()` into that
instead.
I also stumbled across yet another bug,
https://bugs.llvm.org/show_bug.cgi?id=41753, but it seems to be unrelated to
this checker.
Simon Pilgrim [Sun, 5 May 2019 17:10:05 +0000 (17:10 +0000)]
Revert rL359962 : Use DiagRuntimeBehavior for -Wunsequenced to weed out false positives
where either the modification or the other access is unreachable.
........
Try to fix buildbots
Nicolas Lesser [Sun, 5 May 2019 12:35:12 +0000 (12:35 +0000)]
[clang] fixing -ast-print for variadic parameter pack in lambda capture
Summary:
currently for:
```
template<typename ... T>
void f(T... t) {
auto l = [t...]{};
}
```
`clang -ast-print file.cpp`
outputs:
```
template <typename ...T> void f(T ...t) {
auto l = [t] {
}
;
}
```
notice that there is not `...` in the capture list of the lambda. this patch fixes this issue. and add test for it.
Richard Smith [Sat, 4 May 2019 06:46:18 +0000 (06:46 +0000)]
[c++20] Implement tweaked __VA_OPT__ rules from P1042R1:
* __VA_OPT__ is expanded if the *expanded* __VA_ARGS__ is non-empty,
not if the original argument contained no tokens.
* Placemarkers at the start and end of __VA_OPT__ are retained just
long enough to paste them with adjacent ## operators. We never paste
"across" a discarded placemarker.
Richard Trieu [Sat, 4 May 2019 04:22:33 +0000 (04:22 +0000)]
Reduce amount of work ODR hashing does.
When a FunctionProtoType is in the original type in a DecayedType, the decayed
type is a PointerType which points back the original FunctionProtoType. The
visitor for ODRHashing will attempt to process both Type's, doing double work.
By chaining together multiple DecayedType's and FunctionProtoType's, this would
result in 2^N Type's visited only N DecayedType's and N FunctionProtoType's
exsit. Another bug where VisitDecayedType and VisitAdjustedType did
redundant work doubled the work at each level, giving 4^N Type's visited. This
patch removed the double work and detects when a FunctionProtoType decays to
itself to only check the Type once. This lowers the exponential runtime to
linear runtime. Fixes https://bugs.llvm.org/show_bug.cgi?id=41625
Richard Smith [Sat, 4 May 2019 04:00:45 +0000 (04:00 +0000)]
Disallow the operand of __builtin_constant_p from modifying enclosing
state when it's encountered while evaluating a constexpr function.
We attempt to follow GCC trunk's behavior here, but it is somewhat
inscrutible, so our behavior is only approximately the same for now.
Specifically, we only permit modification of objects whose lifetime
began within the operand of the __builtin_constant_p. GCC appears to
have effectively the same restriction, but also some unknown restriction
based on where and how the local state of the constexpr function is
mentioned within the operand (see added testcases).
Richard Smith [Sat, 4 May 2019 00:27:21 +0000 (00:27 +0000)]
[cxx_status] Don't list -fmodules / -fmodules-ts against C++ modules
support; those turn on different modules modes. The real C++ modules
support is behind -std=c++2a like the rest of C++20.
Nicolas Lesser [Sat, 4 May 2019 00:09:00 +0000 (00:09 +0000)]
[clang] adding explicit(bool) from c++2a
this patch adds support for the explicit bool specifier.
Changes:
- The parsing for the explicit(bool) specifier was added in ParseDecl.cpp.
- The storage of the explicit specifier was changed. the explicit specifier was stored as a boolean value in the FunctionDeclBitfields and in the DeclSpec class. now it is stored as a PointerIntPair<Expr*, 2> with a flag and a potential expression in CXXConstructorDecl, CXXDeductionGuideDecl, CXXConversionDecl and in the DeclSpec class.
- Following the AST change, Serialization, ASTMatchers, ASTComparator and ASTPrinter were adapted.
- Template instantiation was adapted to instantiate the potential expressions of the explicit(bool) specifier When instantiating their associated declaration.
- The Add*Candidate functions were adapted, they now take a Boolean indicating if the context allowing explicit constructor or conversion function and this boolean is used to remove invalid overloads that required template instantiation to be detected.
- Test for Semantic and Serialization were added.
This patch is not yet complete. I still need to check that interaction with CTAD and deduction guides is correct. and add more tests for AST operations. But I wanted first feedback.
Perhaps this patch should be spited in smaller patches, but making each patch testable as a standalone may be tricky.
Don Hinton [Fri, 3 May 2019 17:47:29 +0000 (17:47 +0000)]
[CommandLine] Change help output to prefix long options with `--` instead of `-`. NFC . Part 3 of 5
Summary:
By default, `parseCommandLineOptions()` will accept either a
`-` or `--` prefix for long options -- options with names longer than
a single character.
While this change does not affect behavior, it will be helpful with a
subsequent change that requires long options use the `--` prefix.
David Blaikie [Fri, 3 May 2019 00:44:50 +0000 (00:44 +0000)]
SemaOverload: Complete candidates before emitting the error, to ensure diagnostics emitted (or suppressed) during completion don't interfere with the overload notes
Because diagnostics and their notes are not connected at the API level,
if the error message for an overload is emitted, then the overload
candidates are completed - if a diagnostic is emitted during that work,
the notes related to overload candidates would be attached to the latter
diagnostic, not the original error. Sort of worse, if the latter
diagnostic was disabled, the notes are disabled.
Richard Smith [Thu, 2 May 2019 23:21:28 +0000 (23:21 +0000)]
Fix -Wunsequenced false-positives in code controlled by a branch on
__builtin_constant_p.
If the operand of __builtin_constant_p is not constant and has
side-effects, then code controlled by a branch on it is unreachable and
we should not emit runtime behavior warnings in such code.
Leonard Chan [Thu, 2 May 2019 20:38:14 +0000 (20:38 +0000)]
[Attribute/Diagnostics] Print macro if definition is an attribute declaration
If an address_space attribute is defined in a macro, print the macro instead
when diagnosing a warning or error for incompatible pointers with different
address_spaces.
We allow this for all attributes (not just address_space), and for multiple
attributes declared in the same macro.
Nico Weber [Thu, 2 May 2019 19:47:05 +0000 (19:47 +0000)]
Another attempt to fix "could not find clang-check" lit warning in analyzer-less builds
r359717 added clang-check as a dep of check-clang unconditionally
because I had missed lit.local.cfg in test/Tooling.
Instead, only add clang-check to the tools if the analyzer is enabled,
since the build target only exists then, and since all tests using
clang-check are skipped when the analyzer is disabled.
Before this change, when emitting debug info for the
`is_same_v<int, int>` global variable, clang would crash because it
would try to use the template parameter list from the partial
specialization to give parameter names to template arguments. This
doesn't work in general, since a partial specialization can have fewer
arguments than the primary template. Therefore, always use the primary
template. Hypothetically we could try to use the parameter names from
the partial specialization when possible, but it's not clear this really
helps debugging in practice.
When the expression used to initialise 'this' has a pointer type,
check the address space of the pointee type instead of the pointer
type to decide whether an address space cast is required.
It is the pointee type that carries the address space qualifier.
Adapt -fsanitize=function to SANITIZER_NON_UNIQUE_TYPEINFO
This follows up after b7692bc3e9ad2691fc07261904b88fb15f30696b "[UBSan] Fix
isDerivedFromAtOffset on iOS ARM64" fixed the RTTI comparison in
isDerivedFromAtOffset on just one platform and then a25a2c7c9a7e1e328a5bd8274d2d86b1fadc4692 "Always compare C++ typeinfo (based on
libstdc++ implementation)" extended that fix to more platforms.
But there is another RTTI comparison for -fsanitize=function generated in
clang's CodeGenFunction::EmitCall as just a pointer comparison. For
SANITIZER_NON_UNIQUE_TYPEINFO platforms this needs to be extended to also do
string comparison. For that, __ubsan_handle_function_type_mismatch[_abort]
takes the two std::type_info pointers as additional parameters now, checks them
internally for potential equivalence, and returns without reporting failure if
they turn out to be equivalent after all. (NORETURN needed to be dropped from
the _abort variant for that.) Also these functions depend on ABI-specific RTTI
now, so needed to be moved from plain UBSAN_SOURCES (ubsan_handlers.h/cc) to
UBSAN_CXXABI_SOURCES (ubsan_handlers_cxx.h/cc), but as -fsanitize=function is
only supported in C++ mode that's not a problem.
projects / clang / log