Ahmed Bougacha [Thu, 4 Jun 2015 01:43:41 +0000 (01:43 +0000)]
[CodeGen][NEON] Emit constants for "immediate" intrinsic arguments.
On ARM/AArch64, we currently always use EmitScalarExpr for the immediate
builtin arguments, instead of directly emitting the constant. When the
overflow sanitizer is enabled, this generates overflow intrinsics
instead of constants, breaking assumptions in various places.
Instead, use the knowledge of "immediates" to directly emit a constant:
- teach the tablegen backend to emit the "immediate" modifiers
- use those modifiers in the NEON CodeGen, on ARM and AArch64.
Justin Bogner [Thu, 4 Jun 2015 00:30:22 +0000 (00:30 +0000)]
Driver: Don't crash when generating crash reports for *-header inputs
If we crash while handling headers, the crash report mechanism
currently tries to make a string out of a null pointer when it tries
to make up a file extension.
Map *-header input types to reasonable extensions to avoid this.
Russell Gallop [Wed, 3 Jun 2015 15:09:13 +0000 (15:09 +0000)]
[utils] Improvements to check_cfc.py to work better with some build systems.
Recognise options to output dependency files and don't perform checks.
Report input file name when reporting a check failure so it is more obvious in large build logs.
Russell Gallop [Wed, 3 Jun 2015 14:33:57 +0000 (14:33 +0000)]
[utils] Add exact check to check_cfc.py dash_s_no_change.
Files compiled with -via-file-asm should be byte for byte identical. This
change improves the checking on dash_s_no_change to detect non-code
differences. If there is a difference, the check goes on to compare code and
debug to try and be more informative.
Manuel Klimek [Wed, 3 Jun 2015 13:10:41 +0000 (13:10 +0000)]
Allow replacements created from token ranges to specify language options.
The default language options will lead to incorrect replacements in C++
code, for example when trying to replace nested name specifiers ending
in "::".
David Majnemer [Wed, 3 Jun 2015 00:26:35 +0000 (00:26 +0000)]
[Sema] Make the atomic builtins more efficient by reducing volatility
The parameter types and return type do not need to be volatile just
because the pointer type's pointee type is volatile qualified. This is
an unnecessary pessimization.
David Majnemer [Tue, 2 Jun 2015 22:15:12 +0000 (22:15 +0000)]
[MSVC Compatibility] Permit static_cast from void-ptr to function-ptr
The MSVC 2013 and 2015 implementation of std::atomic is specialized for
pointer types. The member functions are implemented using a static_cast
from void-ptr to function-ptr which is not allowed in the standard.
Permit this conversion if -fms-compatibility is present.
[x86-64 ABI] Fix for PR23082: an assertion failure when passing/returning a wrapper union in a full YMM register.
This patch fixes an assertion failure in method
'X86_64ABIInfo::GetByteVectorType'.
Method 'GetByteVectorType' (in TargetInfo.cpp) is responsible
for mapping a QualType 'Ty' (for an argument or return value) to an LLVM IR
type that, according to the ABI, must be passed in a XMM/YMM vector register.
When selecting the IR vector type, method 'GetByteVectorType' always tries to
choose the "best" IR vector type for the 'Ty' in input. In particular, if Ty
is a wrapper structure, it keeps unwrapping it until it finds a vector type VTy.
That VTy is the "preferred IR type".
However, function 'isSingleElementStructure' (used to unwrap structures) does
not know how to look through union types. So, before this patch, if Ty was in
a nest of wrapper structures with at least two union types, we would have
triggered an assertion failure (added at revision 230971).
With this patch, if method 'GetByteVectorType' fails to find the preferred
vector type, we just return a valid (although potentially 'less friendly')
vector type based on the type size. So, rather than asserting on an 'unexpected'
'Ty' in input, we conservatively return vector type <2 x double> if Ty is 16
bytes, or <4 x double> if Ty is 32 bytes.
Sanjay Patel [Tue, 2 Jun 2015 16:55:12 +0000 (16:55 +0000)]
add the -mrecip driver flag and process its options (2nd try)
The first try to land this (r238055) was reverted due to bot failures
caused by the LLVM part of the patch. That was hopefully fixed by r238788,
and the LLVM patch was resubmitted at r238842.
This is the front-end counterpart to D8982.
The -mrecip option interface is based on maintaining compatibility with gcc:
https://gcc.gnu.org/onlinedocs/gcc-4.9.2/gcc/i386-and-x86-64-Options.html#index-mrecip_003dopt-1627
https://gcc.gnu.org/onlinedocs/gcc-4.9.2/gcc/RS_002f6000-and-PowerPC-Options.html#index-mrecip-2289
...while adding more functionality (allowing users to specify the number of refinement steps for each
estimate type).
Manuel Klimek [Tue, 2 Jun 2015 12:01:50 +0000 (12:01 +0000)]
Make vim also output a helpful message in some error cases.
When clang-format encounters a syntax error, it will not format that
line; we're now using the same mechanism we're already using in emacs to
show a helpful error message to the user.
David Majnemer [Mon, 1 Jun 2015 07:41:37 +0000 (07:41 +0000)]
[MS ABI] Be a little more defensive wrt vector types
We probably shouldn't say that all appropriately sized vector types are
intel vector types (i.e. __m128, etc.) as they don't exist for all
architectures. While this is largely academic, it'd save some debugging
if we supported such a platform.
Daniel Jasper [Sun, 31 May 2015 08:51:54 +0000 (08:51 +0000)]
clang-format: NFC. Cleanup after r237895.
Specifically adhere to LLVM Coding Standards (no 'else' after
return/break/continue) and remove yet another implementation of
paren counting. We already have enough of those in the
UnwrappedLineParser.
David Majnemer [Sat, 30 May 2015 09:12:07 +0000 (09:12 +0000)]
[CodeGen] Indirect fields can initialize a union
The first named data member is the field used to default initialize the
union. An IndirectFieldDecl can introduce the first named data member
of a union.
John Brawn [Fri, 29 May 2015 13:10:44 +0000 (13:10 +0000)]
[ARM] Adjust -march checking
getCanonicalArchName can return an empty string for an architecture
that is well-formed but meaningless. Use parseArch to determine if
it's actually valid or not.
Alexander Musman [Fri, 29 May 2015 11:24:32 +0000 (11:24 +0000)]
Fix assertion on C++ attributes in fillAttributedTypeLoc
this fixes http://llvm.org/PR17424
fillAttributedTypeLoc() function is only called with AttributeLists of either
DeclarationChunk (which is used for each type in a declarator being parsed) or
DeclSpec (which captures information about declaration specifiers).
As C++11 attributes actually appertain to declarators, they are moved straight
to the declarator’s attr list in distributeFunctionTypeAttrFromDeclSpec()
function.
'Put them wherever you like' semantics is not supported for C++11 attributes
(but is allowed for GNU attributes, for example). So when we meet an attribute
while parsing the declaration, we cannot be sure if it appertains to either
DeclarationChunk or DeclSpec.
This investigation correlates with the history of changes of SemaType.cpp:
• Asserts in fillAttributedTypeLoc() were added on 3 Mar 2011 in r126986
(http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon- 20110228/039638.html);
• Distributing C++11 attrs to the declarator was added on 14 Jan 2013
in r172504 (http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon- 20130114/071830.html).
Considering all written above I changed asserts in fillAttributedTypeLoc()
to nullptr checks.
This fixes PR17424 and related assertion on
[[gnu::fastcall]] void __stdcall foo();
Diego Novillo [Thu, 28 May 2015 21:30:04 +0000 (21:30 +0000)]
Improve user documentation on profiling.
This clarifies the relationship between instrumentation and sampling based PGO,
code coverage analysis and the different formats supported by sample
profiling.
Chandler Carruth [Thu, 28 May 2015 21:20:14 +0000 (21:20 +0000)]
[omp] Loosen the driver test enough so that overriding the defaults
works well for folks.
This isn't terribly clean (sadly) but after chatting with both Eric and
Richard, nothing cleaner really emerged. The clean way of doing this is
a *lot* of work for extremely little benefit here.
Hans Wennborg [Thu, 28 May 2015 17:44:56 +0000 (17:44 +0000)]
Get the dll storage class right for structors of classes exported/imported via explicit instantiation (PR23667)
This is a follow-up to r238266. It turned out structors are codegened through a different path,
and didn't get the storage class set in EmitGlobalFunctionDefinition.
Daniel Jasper [Thu, 28 May 2015 07:21:50 +0000 (07:21 +0000)]
clang-format: Lower binding strengths created by the [] created by ObjC
method expressions and array literals. They should not bind stronger
than regular parentheses or the braces of braced lists.
Specific test case in JavaScript:
Before:
var aaaaa: List<
SomeThing> = [new SomeThingAAAAAAAAAAAA(), new SomeThingBBBBBBBBB()];
After:
var aaaaa: List<SomeThing> = [
new SomeThingAAAAAAAAAAAA(),
new SomeThingBBBBBBBBB()
];
David Majnemer [Thu, 28 May 2015 04:36:18 +0000 (04:36 +0000)]
[Basic] Define __declspec for cygwin
Cygwin (and MinGW) targets define __declspec to __attribute__ unless
-fms-extensions is specified. It turns out that cygwin headers rely on
the existence of this macro.
Chandler Carruth [Thu, 28 May 2015 01:52:38 +0000 (01:52 +0000)]
[omp] Re-work Clang's handling of -fopenmp and undo r237769.
This isn't an actual revert of r237769, it just restores the behavior of
the Clang driver prior to it while completely re-implementing how that
behavior works.
This also re-does the work of making the default OpenMP runtime
selectable at CMake (or configure) time to work in the way all of our
other such hooks do (config.h, configure and cmake hooks, etc.).
I've re-implemented how we manage the '-fopenmp' flagset in an important
way. Now, the "default" hook just makes '-fopenmp' equivalent to
'-fopenmp=<default>' rather than a separate special beast. Also, there
is an '-fno-openmp' flag which does the obvious thing. Also, the code is
shared between all the places to select a known OpenMP runtime and act
on it.
Finally, and most significantly, I've taught the driver to inspect the
selected runtime when choosing whether to propagate the '-fopenmp' flag
to the frontend in the CC1 commandline. Without this, it isn't possible
to use Clang with libgomp, even if you were happy with the serial,
boring way in which it worked previously (ignoring all #pragmas but
linking in the library to satisfy direct calls into the runtime).
While I'm here, I've gone ahead and sketched out a path for the future
name of LLVM's OpenMP runtime (libomp) and the legacy support for its
current name (libiomp5) in what seems a more reasonable way.
To re-enable LLVM's OpenMP runtime (which I think should wait until the
normal getting started instructions are a reasonable way for falks to
check out, build, and install Clang with the runtime) all that needs to
change is the default string in the CMakeLists.txt and configure.ac
file. No code changes necessary.
I also added a test for the driver's behavior around OpenMP since it was
*completely missing* previously. Makes it unsurprising that we got it
wrong.
Richard Trieu [Thu, 28 May 2015 00:14:02 +0000 (00:14 +0000)]
Update -Winvalid-noreturn to handle destructors better.
When checking if a function is noreturn, consider a codepath to be noreturn if
the path destroys a class and the class destructor, base class destructors, or
member field destructors are marked noreturn.
Daniel Jasper [Wed, 27 May 2015 05:37:40 +0000 (05:37 +0000)]
clang-format: [JS] Fix incorrect detection of ternary expressions.
A definintion like this could not be formatted at all:
constructor({aa}: {
aa?: string, aaaaaaaa?: string, aaaaaaaaaaaaaaa?: boolean,
aaaaaa?: List<string>
}) {
}
Hans Wennborg [Wed, 27 May 2015 00:06:45 +0000 (00:06 +0000)]
clang-cl: Handle dll attributes in explicit class template specialization definitions (PR23667)
Previously, we wouldn't call checkDLLAttribute() after the class template
specialization definition if the class template was already instantiated
by an explicit class template specialization declaration.
David Majnemer [Tue, 26 May 2015 21:54:24 +0000 (21:54 +0000)]
[MS ABI, DebugInfo] Omit the size for model-less pointers-to-members
The representation of a pointer-to-member in the MS ABI is governed by
the layout of the relevant class or if a model has been explicitly
specified. If no model is specified, then an appropriate
"worst-case-scenario" model is implicitly chosen if, and only, if the
pointer-to-member type's representation was needed.
Debug info cannot force a pointer-to-member type to have a
representation so do not try to query the size of such a type unless we
know it is safe to do so.
David Majnemer [Tue, 26 May 2015 21:28:50 +0000 (21:28 +0000)]
[CodeGen] Handle flexible array members containing pointers to members
Types can be classified as being zero-initializable or
non-zero-initializable. We used to classify array types by giving them
the classification of their base element type. However, incomplete
array types are never initialized directly and thus are always
zero-initializable.
Petar Jovanovic [Tue, 26 May 2015 21:07:19 +0000 (21:07 +0000)]
[MIPS] Re-land the change r238200 to fix extension of integer types
Re-land the change r238200, but with modifications in the tests that should
prevent new failures in some environments as reported with the original
change on the mailing list.
Aaron Ballman [Tue, 26 May 2015 19:44:52 +0000 (19:44 +0000)]
__declspec is not a core Clang language extension. Instead, require -fms-extensions or -fborland to enable the language extension.
Note: __declspec is also temporarily enabled when compiling for a CUDA target because there are implementation details relying on __declspec(property) support currently. When those details change, __declspec should be disabled for CUDA targets.
Daniel Jasper [Tue, 26 May 2015 07:26:26 +0000 (07:26 +0000)]
clang-format: Guard the bin-packing in braced lists on BinPackArguments
instead of BinPackParameters. Braced lists are used as constructor
calls in many places and so the bin-packing should follow what is done
for other calls and not what is done for function declarations.
projects / clang / log