Ahmed Bougacha [Tue, 19 Apr 2016 17:54:29 +0000 (17:54 +0000)]
[CodeGen] Widen non-power-of-2 vector HFA base types.
Currently, for the ppc64--gnu and aarch64 ABIs, we recognize:
typedef __attribute__((__ext_vector_type__(3))) float v3f32;
typedef __attribute__((__ext_vector_type__(16))) char v16i8;
struct HFA {
v3f32 a;
v16i8 b;
};
as an HFA. Since the first type encountered is used as the base type,
we pass the HFA as:
[2 x <3 x float>]
Which leads to incorrect IR (relying on padding values) when the
second field is used.
Instead, explicitly widen the vector (after size rounding) in
isHomogeneousAggregate.
If the untied clause is present on a task construct, any thread in the
team can resume the task region after a suspension. Patch adds proper
codegen for untied tasks.
Samuel Benzaquen [Tue, 19 Apr 2016 15:52:56 +0000 (15:52 +0000)]
[ASTMatchers] Do not try to memoize nodes we can't compare.
Summary:
Prevent hasAncestor from comparing nodes that are not supported.
hasDescendant was fixed some time ago to avoid this problem.
I'm applying the same fix to hasAncestor: if any object in the Builder map is
not comparable, skip the cache.
Martin Probst [Tue, 19 Apr 2016 14:55:37 +0000 (14:55 +0000)]
clang-format: [JS] simplify import/export.
Summary:
Change `import` and `export` parsing to special case the renaming
syntax (`import x, {y as bar} ...`, `export {x}`) and otherwise just
parse a regular structural element.
This simplifies the code a bit and should be more correct - it's easier
to recognise the specific import syntax than to recognise arbitrary
expressions and declarations.
If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks.
Warn if function or variable cannot be implicitly instantiated
With this patch compiler emits warning if it tries to make implicit instantiation
of a template but cannot find the template definition. The warning can be suppressed
by explicit instantiation declaration or by command line options
-Wundefined-var-template and -Wundefined-func-template. The implementation follows
the discussion of http://reviews.llvm.org/D12326.
Richard Smith [Mon, 18 Apr 2016 23:12:59 +0000 (23:12 +0000)]
Mark -Xclang as being a compilation-only option so that the clang driver
doesn't warn if it's passed to a link action. This matches the behavior for
most other compilation-only options (including things like -f flags), and is
necessary to suppress warnings on systems like cmake that always pass all
compile flags to the linker.
The builtins already exist in LLVM, but are not exposed to the C/C++
programmers. This patch adds all the information about the builtins
needed for clang, as well as a test for all available intrinsics.
When an explicit specialization of a function template, a member function template or a member function of a class template is declared, clang first implicitly instantiates the declaration of a specialization from the templated-entity being explicitly specialized (since their signatures must be the same) and then links the explicit specialization being declared as a redeclaration of the aforementioned specialization.
The problem was that when clang 'implicitly instantiates' the initial specialization, it marks the corresponding FunctionDecl as deleted if the corresponding templated-entity was deleted, rather than waiting to see whether the explicit specialization being declared provides a non-deleted body. (The eager marking of delete has advantages during overload resolution I suppose, where we don't have to try and instantiate a definition of the function to see if it is deleted).
The present fix entails recognizing that when clang knows that an explicit specialization is being declared (for whichever templated-entity), the prior implicit instantiation should not inherit the 'deleted' status, and so we reset it to false.
I suppose an alternative fix (amongst others) could consider creating a new context (ExplicitSpecializationDeclarationSubstitution or some such) that is checked during template-argument-deduction and final substitution, and avoid inheriting the deleted status during declaration substitution. But while conceptually cleaner, that would be a slightly more involved change (as could be some of the other alternatives: such as avoid tagging implicit specializations as deleted, and check their primary templates for the deleted status where needed), and so I chose a different path. Hopefully it'll prove to not be a bad choice.
DebugInfo: Make DICompositeTypes distinct most of the time
Since elements of most kinds of DICompositeType have back references,
most are involved in uniquing cycles. Except via the ODR 'identifier:'
field, which doesn't care about the storage type (see r266549),
they have no hope of being uniqued.
Distinct nodes are far more efficient, so use them for most kinds of
DICompositeType definitions (i.e., when DIType::isForwardDecl is false).
The exceptions:
- DW_TAG_array_type, since their elements never have back-references
and they never have ODR 'identifier:' fields;
- DW_TAG_enumeration_type when there is no ODR 'identifier:' field,
since their elements usually don't have back-references.
This breaks the last major uniquing cycle I'm aware of in the debug info
graph. The impact won't be enormous for C++ because references to
ODR-uniqued nodes still use string-based DITypeRefs; but this should
prevent a regression in C++ when we drop the string-based references.
This wouldn't have been reasonable until r266549, when composite types
stopped relying on being uniqued by structural equivalence to prevent
blow-ups at LTO time.
Richard Smith [Sat, 16 Apr 2016 00:46:26 +0000 (00:46 +0000)]
[modules] Don't expose *intrin.h headers that cannot be included standalone as
separate modules. These cause build breakage with -fmodules-local-submodule-visibility.
[CUDA] Raise an error if the CUDA install can't be found.
Summary:
Without this change, we silently proceed on without including
__clang_cuda_runtime_wrapper.h. This leads to very strange behavior --
you say you're compiling CUDA code, but e.g. __device__ is not defined!
Revert 266186 as it breaks anything that includes type_traits on some platforms
Since this patch provided support for the __float128 type but disabled it
on all platforms by default, some platforms can't compile type_traits with
-std=gnu++11 since there is a specialization with __float128.
This reverts the patch until D19125 is approved (i.e. we know which platforms
need this support enabled).
[MSVC Compat] Implementation of __unaligned (MS extension) as a type qualifier
This patch implements __unaligned as a type qualifier; before that, it was
modeled as an attribute. Proper mangling of __unaligned is implemented as well.
Some OpenCL code/tests are tangenially affected, as they relied on existing
number and sizes of type qualifiers.
Hans Wennborg [Fri, 15 Apr 2016 00:24:15 +0000 (00:24 +0000)]
clang-cl: Make /link accept an optional joined argument.
For example, "cl.exe a.c /linkfoo bar" is a valid invocation and
forwards "foo" and "bar" to link.exe. This makes clang-cl handle
that kind of invocation.
Richard Smith [Thu, 14 Apr 2016 00:50:18 +0000 (00:50 +0000)]
[modules] Store the location of the lexical update record in a decl update as
an offset from the current record rather than as an absolute bit number. This
gives a minor .pcm file size reduction.
Richard Smith [Thu, 14 Apr 2016 00:29:55 +0000 (00:29 +0000)]
[modules] Remove CXX_BASE_SPECIFIERS_OFFSETS table. Instead of storing an ID of
a table entry in the corresponding decl, store an offset from the current
record to the relevant CXX_BASE_SPECIFIERS record. This results in fewer
indirections and a minor .pcm file size reduction.
ObjC kindof: order the methods in global pool relative to availability.
r265877 tries to put methods that are deprecated or unavailable to the
front of the global pool to emit diagnostics, but it breaks some of
our existing codes that depend on choosing a certain method for id
lookup.
This commit orders the methods with the same declaration with respect
to the availability, but do not order methods with different declaration.
[CodeGen] Avoid ctor/dtor boilerplate with some C++11
Non-owning pointers that cache LLVM types and constants can use
'nullptr' default member initializers so that we don't need to mention
them in the constructor initializer list.
Owning pointers should use std::unique_ptr so that we don't need to
manually delete them in the destructor. They also don't need to be
mentioned in the constructor at that point.
Richard Smith [Wed, 13 Apr 2016 21:57:08 +0000 (21:57 +0000)]
[modules] Remove CXX_CTOR_INITIALIZERS_OFFSETS table. Instead of storing an ID
of a table entry in the corresponding decl, store an offset from the current
record to the relevant CXX_CTOR_INITIALIZERS record. This results in fewer
indirections and a minor .pcm file size reduction.
[SemaObjC] Properly handle mix between type arguments and protocols.
Under certain conditions clang currently fails to properly diagnostic ObjectC
parameter list when type args and protocols are mixed in the same list. This
happens when the first item in the parameter list is a (1) protocol, (2)
unknown type or (3) a list of protocols/unknown types up to the first type
argument. Fix the problem to report the proper error, example:
[CrashReproducer] Setup 'use-external-names' in YAML files.
Hide the real paths when rebuilding from VFS by setting up the crash
reproducer to use 'use-external-names' = false. This way we avoid
module redifinition errors and consistently use the same paths against
all modules.
With this change on Darwin we are able to simulate a crash for a simple
application using "Foundation/Foundation.h" (which relies on a bunch of
different frameworks and headers) and successfully rebuild all the
modules by relying solely at the VFS overlay.
Chris Bieneman [Wed, 13 Apr 2016 17:12:56 +0000 (17:12 +0000)]
[OrderFiles] Don't allow lit to run dtrace multithreaded
Dtrace is implemented to try and minimize performance impact on the process being traced. This results in dtrace dropping samples if it is taking too many CPU resources. Multi-threading dtrace increases the sample drop rate dramatically.
Tim Northover [Wed, 13 Apr 2016 17:08:55 +0000 (17:08 +0000)]
AArch64: allow 64-bit access to sysregs.
Although all the registers are actually 32-bits, I think we have to assume the
high 32-bits could be RES0 (the ARM ARM is unclear). If so, reading as a 32-bit
register can require extra zero extension operations.
Tim Northover [Wed, 13 Apr 2016 17:08:51 +0000 (17:08 +0000)]
ARM: make Darwin's "-arch armv7em" default to hard-float.
We've already paid the price for separate "armv7m" and "armv7em" slices
(support in other tools), it's silly to make them identical other than the
default CPU.
Add functions declared in ctype.h to builtin function database. All functions are annotated with nothrow and pure attribute, which enables better optimization.
Add AST Matchers for CXXConstructorDecl::isDelegatingConstructor and CXXMethodDecl::isUserProvided.
Summary: Added two AST matchers: isDelegatingConstructor for CXXConstructorDecl::IsDelegatingConstructor; and isUserProvided corresponding to CXXMethodDecl::isUserProvided.
This patch corresponds to review:
http://reviews.llvm.org/D15120
It adds support for the __float128 keyword, literals and a target feature to
enable it. This support is disabled by default on all targets and any target
that has support for this type is free to add it.
Based on feedback that I've received from target maintainers, this appears to
be the right thing for most targets. I have not heard from the maintainers of
X86 which I believe supports this type. I will subsequently investigate the
impact of enabling this on X86.