Re-commit r363191 "[MS] Pretend constexpr variable template specializations are inline"
While the next Visual Studio update (16.3) will fix this issue, that hasn't
shipped yet. Until then Clang wouldn't work with MSVC's headers which seems
unfortunate. Let's keep this in until VS 16.3 ships. (See also PR42843.)
> Fixes link errors with clang and the latest Visual C++ 14.21.27702
> headers, which was reported as PR42027.
>
> I chose to intentionally make these things linkonce_odr, i.e.
> discardable, so that we don't emit definitions of these things in every
> translation unit that includes STL headers.
>
> Note that this is *not* what MSVC does: MSVC has not yet implemented C++
> DR2387, so they emit fully specialized constexpr variable templates with
> static / internal linkage.
>
> Reviewers: rsmith
>
> Differential Revision: https://reviews.llvm.org/D63175
------------------------------------------------------------------------
[RISCV] Set MaxAtomicInlineWidth and MaxAtomicPromoteWidth for RV32/RV64 targets with atomics
Summary: This ensures that libcalls aren't generated when the target supports atomics. Atomics aren't in the base RV32I/RV64I instruction sets, so MaxAtomicInlineWidth and MaxAtomicPromoteWidth are set only when the atomics extension is being targeted. This must be done in setMaxAtomicWidth, as this should be done after handleTargetFeatures has been called.
PR42513: Enter the proper DeclContext before substituting into an
default template argument expression.
We already did this for type template parameters and template template
parameters, but apparently forgot to do so for non-type template
parameters. This causes the substituted default argument expression to
be substituted in the proper context, and in particular to properly mark
its subexpressions as odr-used.
------------------------------------------------------------------------
PR40674: fix assertion failure if a structured binding declaration has a
tuple-like decomposition that produces value-dependent reference
bindings.
------------------------------------------------------------------------
Until recently, Python_ADDITIONAL_VERSIONS was used to limit LLVM's
Python support to 2.7. Now that both LLVM and LLDB both support Python
3, there's no longer a need to put an arbitrary limit on this.
However, instead of removing the variable, r365692 expanded the list,
which has the (presumably unintentional) side-effect of expression
preference for Python 3.
Instead, as Michal proposed in the original code review, we should just
not set the list at all, and let CMake pick whatever Python interpreter
you have in your path.
This patch removes the Python_ADDITIONAL_VERSIONS variable in llvm,
clang and lld. I've also updated the docs with the default behavior and
how to force a different Python version to be used.
[OpenCL] Fix addr space deduction for pointers/references to arrays.
Rewrite the logic for detecting if we are deducing addr space of
a pointee type to take into account special logic for arrays. For
pointers/references to arrays we can have any number of parentheses
expressions as well as nested pointers.
[Sema] Implement DR2386 for C++17 structured binding
Allow implementations to provide complete definitions of
std::tuple_size<T>, but to omit the 'value' member to signal that T is
not tuple-like. The Microsoft standard library implements
std::tuple_size<const T> this way.
If the value member exists, clang still validates that it is an ICE, but
if it does not, then the type is considered to not be tuple-like.
Hans Wennborg [Tue, 20 Aug 2019 07:21:14 +0000 (07:21 +0000)]
Revert r365860 for PR42966 (with a tweak to the test case for r365862)
> Author: maskray
> Date: Thu Jul 11 19:01:51 2019
> New Revision: 365860
>
> URL: http://llvm.org/viewvc/llvm-project?rev=365860&view=rev
> Log:
> [Driver] Refactor interaction between -f(no-)?omit-frame-pointer and -m(no-)?omit-leaf-frame-pointer
>
> Use a tri-state enum to represent shouldUseFramePointer() and
> shouldUseLeafFramePointer().
>
> This simplifies the logic and fixes PR9825:
> -fno-omit-frame-pointer doesn't imply -mno-omit-leaf-frame-pointer.
>
> and PR24003:
> /Oy- /O2 should not omit leaf frame pointer: this matches MSVC x86-32.
> (/Oy- is a no-op on MSVC x86-64.)
>
> and:
> when CC1 option -mdisable-fp-elim if absent, -momit-leaf-frame-pointer
> can also be omitted.
>
> The new behavior matches GCC:
> -fomit-frame-pointer wins over -mno-omit-leaf-frame-pointer
> -fno-omit-frame-pointer loses out to -momit-leaf-frame-pointer
>
> The behavior makes lots of sense. We have 4 states:
>
> - 00) leaf retained, non-leaf retained
> - 01) leaf retained, non-leaf omitted (this is invalid)
> - 10) leaf omitted, non-leaf retained (what -momit-leaf-frame-pointer was designed for)
> - 11) leaf omitted, non-leaf omitted
>
> "omit" options taking precedence over "no-omit" options is the only way
> to make 3 valid states representable with -f(no-)?omit-frame-pointer and
> -m(no-)?omit-leaf-pointer.
>
> Reviewed By: ychen
>
> Differential Revision: https://reviews.llvm.org/D64294
Fix handling of class member access into a vector type.
When handling a member access into a non-class, non-ObjC-object type, we
would perform a lookup into the surrounding scope as if for an
unqualified lookup. If the member access was followed by a '<' and this
lookup (or the typo-correction for it) found a template name, we'd treat
the member access as naming that template.
Now we treat such accesses as never naming a template if the type of the
object expression is of vector type, so that vector component accesses
are never misinterpreted as naming something else. This is not entirely
correct, since it is in fact valid to name a template from the enclosing
scope in this context, when invoking a pseudo-destructor for the vector
type via an alias template, but that's very much a corner case, and this
change leaves that case only as broken as the corresponding case for
Objective-C types is.
This incidentally adds support for dr2292, which permits a 'template'
keyword at the start of a member access naming a pseudo-destructor.
------------------------------------------------------------------------
[OpenCL] Ignore parentheses for sampler initialization
The sampler handling logic in SemaInit.cpp would inadvertently treat
parentheses around sampler arguments as an implicit cast, leading to
an unreachable "can't implicitly cast lvalue to rvalue with
this cast kind". Fix by ignoring parentheses once we are in the
sampler initializer case.
Delay diagnosing asm constraints that require immediates until after inlining
Summary:
An inline asm call may result in an immediate input value after inlining.
Therefore, don't emit a diagnostic here if the input isn't an immediate.
Summary:
This adds the 'f' inline assembly constraint, as supported by GCC. An
'f'-constrained operand is passed in a floating point register. Exactly
which kind of floating-point register (32-bit or 64-bit) is decided
based on the operand type and the available standard extensions (-f and
-d, respectively).
This patch adds support in both the clang frontend, and LLVM itself.
Hans Wennborg [Mon, 5 Aug 2019 07:27:39 +0000 (07:27 +0000)]
Merging r367520:
------------------------------------------------------------------------
r367520 | hans | 2019-08-01 10:01:09 +0200 (Thu, 01 Aug 2019) | 15 lines
Delay emitting dllexport explicitly defaulted members until the class is fully parsed (PR40006)
This is similar to r245139, but that only addressed dllexported classes.
It was still possible to run into the same problem with dllexported
members in an otherwise normal class (see bug). This uses the same
strategy to fix: delay defining the method until the whole class has
been parsed.
(The easiest way to see the ordering problem is in
Parser::ParseCXXMemberSpecification(): it calls
ParseLexedMemberInitializers() *after* ActOnFinishCXXMemberDecls(),
which was trying to define the dllexport method. Now we delay it to
ActOnFinishCXXNonNestedClass() which is called after both of those.)
[Preprocessor] Always discard body of #define if we failed to parse it
Summary:
Preivously we would only discard it if we failed to parse parameter lists.
If we do not consume the body, parser sees tokens inside directive. In
turn, this leads to spurious diagnostics and a crash in TokenBuffer, see
the added tests.
[Fix] Customize warnings for missing built-in types
If we detect a built-in declaration for which we cannot derive a type
matching the pattern in the Builtins.def file, we currently emit a
warning that the respective header is needed. However, this is not
necessarily the behavior we want as it has no connection to the location
of the declaration (which can actually be in the header in question).
Instead, this warning is generated
- if we could not build the type for the pattern on file (for some
reason). Here we should make the reason explicit. The actual problem
is otherwise circumvented as the warning is misleading, see [0] for
an example.
- if we could not build the type for the pattern because we do not
have a type on record, possible since D55483, we should not emit any
warning. See [1] for a legitimate problem.
This patch address both cases. For the "setjmp" family a new warning is
introduced and for built-ins without type on record, so far
"pthread_create", we do not emit the warning anymore.
make check-all currently fails on x86_64-pc-solaris2.11 when building with GCC 9:
Undefined first referenced
symbol in file
_ZN11__sanitizer14internal_lseekEimi SANITIZER_TEST_OBJECTS.sanitizer_libc_test.cc.i386.o
_ZN11__sanitizer23MapWritableFileToMemoryEPvmim SANITIZER_TEST_OBJECTS.sanitizer_libc_test.cc.i386.o
ld: fatal: symbol referencing errors
clang-9: error: linker command failed with exit code 1 (use -v to see invocation)
make[3]: *** [projects/compiler-rt/lib/sanitizer_common/tests/CMakeFiles/TSanitizer-i386-Test.dir/build.make:92: projects/compiler-rt/lib/sanitizer_common/tests/Sanitizer-i386-Test] Error 1
While e.g. __sanitizer::internal_lseek is defined in sanitizer_solaris.cc, g++ 9
predefines _FILE_OFFSET_BITS=64 while clang++ currently does not.
This patch resolves this inconsistency by following the gcc lead, which allows
make check-all to finish successfully.
There's one caveat: gcc defines _LARGEFILE_SOURCE and _LARGEFILE64_SOURCE for C++ only, while clang has long been doing it for
all languages. I'd like to keep it this way because those macros do is to make
declarations of fseek/ftello (_LARGEFILE_SOURCE) resp. the 64-bit versions
of largefile functions (*64 with _LARGEFILE64_SOURCE) visible additionally.
However, _FILE_OFFSET_BITS=64 changes all affected functions to be largefile-aware.
I'd like to restrict this to C++, just like gcc does.
To avoid a similar inconsistence with host compilers that don't predefine _FILE_OFFSET_BITS=64
(e.g. clang < 9, gcc < 9), this needs a compantion patch https://reviews.llvm.org/D64483.
[COFF][ARM64] Reorder handling of aarch64 MSVC builtins
In `CodeGenFunction::EmitAArch64BuiltinExpr()`, bulk move all of the aarch64 MSVC-builtin cases to an earlier point in the function (the `// Handle non-overloaded intrinsics first` switch block) in order to avoid an unreachable in `GetNeonType()`. The NEON type-overloading logic is not appropriate for the Windows builtins.
The Arm C Language Extensions for SVE document specifies that
__ARM_FEATURE_SVE should be set when the compiler supports SVE and
implements all the extensions described in the document.
This is currently not yet the case, so the feature should be disabled
until the compiler can provide all the extensions as described.
Add an explicit triple for the test which is pattern matching overly
aggressively.
------------------------------------------------------------------------
[Sema] Fix -Wuninitialized for struct assignment from GNU C statement expression
Summary:
Do not automatically report self references of structs in statement expression
as warnings. Instead wait for uninitialized cfg analysis.
https://bugs.llvm.org/show_bug.cgi?id=42604
This seems to be an old vestage of a previous implementation of getting
the default calling convention, and everything is now using
CXXABI/ASTContext's getDefaultCallingConvention. Remove it, since it
isn't doing anything.
------------------------------------------------------------------------
Following up on the buildbot failures, this commits relaxes some tests:
instead of checking for specific IR output, it now ensures that the
underlying issue (the crash), and only that, doesn't happen.
------------------------------------------------------------------------
Summary:
After r345152 cached completions started adding namespaces after
nested name specifiers, e.g. in `some_name::^`
The CCC_Symbol indicates the completed item cannot be a namespace (it is
described as being "a type, a function or a variable" in the comments).
Therefore, 'nested specifier' completions should only be added from cache
when the context is CCC_SymbolOrNewName (which roughly seems to indicate
that a nested name specifier is allowed).
[OPENMP]Fix sharing of threadprivate variables with TLS support.
If the threadprivate variable is used in the copyin clause on inner
parallel directive with TLS support, we capture this variable in all
outer OpenMP scopes. It leads to the fact that in all scopes we're
working with the original variable, not the threadprivate copies.
------------------------------------------------------------------------
[OPENMP]Provide correct data sharing attributes for loop control
variables.
Loop control variables are private in loop-based constructs and we shall
take this into account when generate the code for inner constructs.
Currently, those variables are reported as shared in many cases. Moved
the analysis of the data-sharing attributes of the loop control variable
to an early semantic stage to correctly handle their attributes.
------------------------------------------------------------------------
The RISC-V hard float calling convention requires the frontend to:
* Detect cases where, once "flattened", a struct can be passed using
int+fp or fp+fp registers under the hard float ABI and coerce to the
appropriate type(s)
* Track usage of GPRs and FPRs in order to gate the above, and to
determine when signext/zeroext attributes must be added to integer
scalars
This patch attempts to do this in compliance with the documented ABI,
and uses ABIArgInfo::CoerceAndExpand in order to do this. @rjmccall, as
author of that code I've tagged you as reviewer for initial feedback on
my usage.
Note that a previous version of the ABI indicated that when passing an
int+fp struct using a GPR+FPR, the int would need to be sign or
zero-extended appropriately. GCC never did this and the ABI was changed,
which makes life easier as ABIArgInfo::CoerceAndExpand can't currently
handle sign/zero-extension attributes.
Re-landed after backing out 366450 due to missed hunks.
Revert r366422: [OpenCL] Improve destructor support in C++ for OpenCL
Reason: this commit causes crashes in the clang compiler when building
LLVM Support with libc++, see https://bugs.llvm.org/show_bug.cgi?id=42665
for details.
------------------------------------------------------------------------
Marco Antognini [Thu, 18 Jul 2019 10:04:18 +0000 (10:04 +0000)]
[OpenCL] Improve destructor support in C++ for OpenCL
Summary:
This patch does mainly three things:
1. It fixes a false positive error detection in Sema that is similar to
D62156. The error happens when explicitly calling an overloaded
destructor for different address spaces.
2. It selects the correct destructor when multiple overloads for
address spaces are available.
3. It inserts the expected address space cast when invoking a
destructor, if needed, and therefore fixes a crash due to the unmet
assertion in llvm::CastInst::Create.
The following is a reproducer of the three issues:
kernel void foo() {
myGlobal.~MyType(); // 1 and 2.
// 1. error: cannot initialize object parameter of type
// '__generic MyType' with an expression of type '__constant MyType'
// 2. error: no matching member function for call to '~MyType'
}
kernel void bar() {
// 3. The implicit call to the destructor crashes due to:
// Assertion `castIsValid(op, S, Ty) && "Invalid cast!"' failed.
// in llvm::CastInst::Create.
MyType myLocal;
}
The added test depends on D62413 and covers a few more things than the
above reproducer.
[OpenCL][PR42033] Fix addr space deduction with template parameters
If dependent types appear in pointers or references we allow addr
space deduction because the addr space in template argument will
belong to the pointee and not the pointer or reference itself.
We also don't diagnose addr space on a function return type after
template instantiation. If any addr space for the return type was
provided on a template parameter this will be diagnosed during the
parsing of template definition.
This will let us instrument globals during initialization. This required
making the new PM pass a module pass, which should still provide access to
analyses via the ModuleAnalysisManager.
Renamed and changed the wording of warn_cconv_ignored
As discussed in D64780 the wording of this warning message is being
changed to say 'is not supported' instead of 'ignored', and the
diag ID itself is being changed to warn_cconv_not_supported.
Mike Rice [Wed, 17 Jul 2019 15:18:45 +0000 (15:18 +0000)]
[OPENMP]Fix crash in LoopCounterRefChecker when MemberExpr is not Var or Field
checkDecl is only valid for VarDecls or FieldDecls, since getCanonicalDecl
expects only these. Prevent other Decl kinds (such as CXXMethodDecls and
EnumConstantDecls) from entering and asserting.
Summary:
The structural equivalence check reported false eq between lambda classes
with different parameters in their call signature.
The solution is to check the methods for equality too in case of lambda
classes.
[ASTImporter] Fix LLDB lookup in transparent ctx and with ext src
Summary:
With LLDB we use localUncachedLookup(), however, that fails to find
Decls when a transparent context is involved and the given DC has
external lexical storage. The solution is to use noload_lookup, which
works well with transparent contexts. But, we cannot use only the
noload_lookup since the slow case of localUncachedLookup is still needed
in some other cases.
These other cases are handled in ASTImporterLookupTable, but we cannot
use that with LLDB since that traverses through the AST which initiates
the load of external decls again via DC::decls().
We must avoid loading external decls during the import becuase
ExternalASTSource is implemented with ASTImporter, so external loads
during import results in uncontrolled and faulty import.
Summary:
Thread local variables are placed inside a `.tdata` segment. Their symbols are
offsets from the start of the segment. The address of a thread local variable
is computed as `__tls_base` + the offset from the start of the segment.
`.tdata` segment is a passive segment and `memory.init` is used once per thread
to initialize the thread local storage.
`__tls_base` is a wasm global. Since each thread has its own wasm instance,
it is effectively thread local. Currently, `__tls_base` must be initialized
at thread startup, and so cannot be used with dynamic libraries.
`__tls_base` is to be initialized with a new linker-synthesized function,
`__wasm_init_tls`, which takes as an argument a block of memory to use as the
storage for thread locals. It then initializes the block of memory and sets
`__tls_base`. As `__wasm_init_tls` will handle the memory initialization,
the memory does not have to be zeroed.
To help allocating memory for thread-local storage, a new compiler intrinsic
is introduced: `__builtin_wasm_tls_size()`. This instrinsic function returns
the size of the thread-local storage for the current function.
The expected usage is to run something like the following upon thread startup:
Ben Hamilton [Tue, 16 Jul 2019 21:29:40 +0000 (21:29 +0000)]
[clang-format] Don't detect call to ObjC class method as C++11 attribute specifier
Summary:
Previously, clang-format detected something like the following as a C++11 attribute specifier.
@[[NSArray class]]
instead of an array with an Objective-C method call inside. In general, when the attribute specifier checking runs, if it sees 2 identifiers in a row, it decides that the square brackets represent an Objective-C method call. However, here, `class` is tokenized as a keyword instead of an identifier, so this check fails.
To fix this, the attribute specifier first checks whether the first square bracket has an "@" before it. If it does, then that square bracket is not the start of a attribute specifier because it is an Objective-C array literal. (The assumption is that @[[.*]] is not valid C/C++.)
Yonghong Song [Tue, 16 Jul 2019 17:24:33 +0000 (17:24 +0000)]
fix unnamed fiefield issue and add tests for __builtin_preserve_access_index intrinsic
The original commit is r366076. It is temporarily reverted (r366155)
due to test failure. This resubmit makes test more robust by accepting
regex instead of hardcoded names/references in several places.
This is a followup patch for https://reviews.llvm.org/D61809.
Handle unnamed bitfield properly and add more test cases.
Fixed the unnamed bitfield issue. The unnamed bitfield is ignored
by debug info, so we need to ignore such a struct/union member
when we try to get the member index in the debug info.
D61809 contains two test cases but not enough as it does
not checking generated IRs in the fine grain level, and also
it does not have semantics checking tests.
This patch added unit tests for both code gen and semantics checking for
the new intrinsic.
Signed-off-by: Yonghong Song <yhs@fb.com>
git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@366231 91177308-0d34-0410-b5e6-96231b3b80d8
[Driver] Don't pass --dynamic-linker to ld on Solaris
I noticed that clang currently passes --dynamic-linker to ld. This has been the case
since Solaris 11 support was added initially back in 2012 by David Chisnall (r150580).
I couldn't find any patch submission, let alone a justification, for this, and it seems
completely useless: --dynamic-linker is a gld compatibility form of the option, the
native option being -I. First of all, however, the dynamic linker passed is simply the
default, so there's no reason at all to specify it in the first place.
This patch removes passing the option and adjusts the affected testcase accordingly.
Tested on x86_64-pc-solaris2.11 and sparcv9-sun-solaris2.11.
[AArch64] Implement __jcvt intrinsic from Armv8.3-A
The jcvt intrinsic defined in ACLE [1] is available when ARM_FEATURE_JCVT is defined.
This change introduces the AArch64 intrinsic, wires it up to the instruction and a new clang builtin function.
The __ARM_FEATURE_JCVT macro is now defined when an Armv8.3-A or higher target is used.
I've implemented the target detection logic in Clang so that this feature is enabled for architectures from armv8.3-a onwards (so -march=armv8.4-a also enables this, for example).
* Bump the function_type_mismatch handler version, as its signature has changed.
* The function_type_mismatch handler can return successfully now, so
SanitizerKind::Function must be AlwaysRecoverable (like for
SanitizerKind::Vptr).
* But the minimal runtime would still unconditionally treat a call to the
function_type_mismatch handler as failure, so disallow -fsanitize=function in
combination with -fsanitize-minimal-runtime (like it was already done for
-fsanitize=vptr).
Bob Haarman [Tue, 16 Jul 2019 01:35:49 +0000 (01:35 +0000)]
reland "add -fthinlto-index= option to clang-cl"
Summary:
This is a reland of r366146, adding in the previously missing '--'
flag that prevents filenames from being interpreted as flags.
Original description:
This adds a -fthinlto-index= option to clang-cl, which allows it to
be used to drive ThinLTO backend passes. This allows clang-cl to be
used for distributed ThinLTO.
Peter Wu [Tue, 16 Jul 2019 01:13:36 +0000 (01:13 +0000)]
[Sema] Suppress additional warnings for C's zero initializer
Summary:
D28148 relaxed some checks for assigning { 0 } to a structure for all C
standards, but it failed to handle structures with non-integer
subobjects. Relax -Wmissing-braces checks for such structures, and add
some additional tests.
Bob Haarman [Mon, 15 Jul 2019 22:50:04 +0000 (22:50 +0000)]
add -fthinlto-index= option to clang-cl
Summary:
This adds a -fthinlto-index= option to clang-cl, which allows it to
be used to drive ThinLTO backend passes. This allows clang-cl to be
used for distributed ThinLTO.
The default value for this option (UseMCJITReplacement) is already false, and
OrcMCJITReplacement is going to have deprecation warnings attached in LLVM 9.0.
Removing this call removes a spurious warning.
projects / clang / log