Alp Toker [Wed, 16 Jul 2014 16:48:33 +0000 (16:48 +0000)]
Make clang's rewrite engine a core feature
The rewrite facility's footprint is small so it's not worth going to these
lengths to support disabling at configure time, particularly since key compiler
features now depend on it.
Meanwhile the Objective-C rewriters have been moved under the
ENABLE_CLANG_ARCMT umbrella for now as they're comparatively heavy and still
potentially worth excluding from lightweight builds.
Tests are now passing with any combination of feature flags. The flags
historically haven't been tested by LLVM's build servers so caveat emptor.
Objective-C. Introducing __attribute__((objc_runtime_name("runtimename"))
to be applied to class or protocols. This will direct IRGen
for Objective-C metadata to use the new name in various places
where class and protocol names are needed.
rdar:// 17631257
Alp Toker [Wed, 16 Jul 2014 15:12:48 +0000 (15:12 +0000)]
Revert "clang/test/Driver/crash-report.c: This requires rewriter for -frewrite-includes. [PR20321]"
We've decided to make the core rewriter class and PP rewriters mandatory.
They're only a few hundred lines of code in total and not worth supporting as a
distinct build configuration, especially since doing so disables key compiler
features.
This reverts commit r213150.
Revert "clang/test: Introduce the feature "rewriter" for --enable-clang-rewriter."
This reverts commit r213148.
Revert "Move clang/test/Frontend/rewrite-*.c to clang/test/Frontend/Rewriter/"
Daniel Sanders [Wed, 16 Jul 2014 11:52:23 +0000 (11:52 +0000)]
[mips] Correct the invocation of GAS in several cases.
Summary:
As a result of this patch, assembling an empty file with GCC and Clang (using
GAS as the assembler) now produces identical objects.
-mfp32/-mfpxx/-mfp64 now form a trinity of options. -mfpxx is the default
when the triple vendor is 'img' or 'mti', the ABI is O32, and the CPU is
between mips2 and mips32r2/mips64r2 (inclusive).
-mno-shared is always given to the assembler to match the effect of
-mabicalls (currently unimplemented but Clang acts as if it is given).
Similarly, -call_nonpic is always given to match the effect of -mplt (also
unimplemented and acts as if given) except when the ABI is 64 in which case
-mplt has no effect so -KPIC is given instead.
-mhard-float/-msoft-float are now passed on.
-modd-spreg/-mno-odd-spreg are now passed on.
-mno-mips16 is correctly passed on. The assembler option is -no-mips16 not
-mno-mips16
David Majnemer [Wed, 16 Jul 2014 07:16:58 +0000 (07:16 +0000)]
MS ABI: Up the required alignment after inserting padding between vbases
We would correctly insert sufficiently aligned padding between vbases
when our leading base was empty, however we would neglect to increase
the required alignment of the most derived class.
David Majnemer [Wed, 16 Jul 2014 06:04:00 +0000 (06:04 +0000)]
AST: Cleanup RecordLayoutBuilder
No functionality changed, just some cleanups:
- Switch some loops to range-based for.
- Name some iterators with a more creative name than "I".
- Reduce dependence on auto. Does RD->bases() give you a list of
CXXBaseSpecifiers or CXXRecordDecls? It's more clear to just say which
upfront.
Recognize additional cases, when '::' is mistyped as ':'.
This is a fix to RP18587 - colons have too much protection in member-declarations
Review is tracked by http://reviews.llvm.org/D3653.
This is an attempt to recommit the fix, initially committed as r212957 but then
reverted in r212965 as it broke self-build. In the updated patch ParseDirectDeclarator
turns on colon protection in for context as well.
Driver: bifurcate extended and basic MSC versioning
This restores the original behaviour of -fmsc-version. The older option
remains as a mechanism for specifying the basic version information. A
secondary option, -fms-compatibility-version permits the user to specify an
extended version to the driver.
The new version takes the value as a dot-separated value rather than the
major * 100 + minor format that -fmsc-version format. This makes it easier to
specify the value as well as a more flexible manner for specifying the value.
Specifying both values is considered an error.
The older parameter is left solely as a driver option, which is normalised into
the newer parameter. This allows us to retain a single code path in the
compiler itself whilst preserving the semantics of the old parameter as well as
avoid having to determine which of two formats are being used by the invocation.
The test changes are due to the fact that the compiler no longer supports the
old option, and is a direct conversion to the new option.
Handle diagnostic warnings in Frontend diagnostic handler.
Clang uses a diagnostic handler to grab diagnostic messages so it can print them
with the line of source code they refer to. This patch extends this to handle
diagnostic warnings that were added to llvm to produce a warning when
loop vectorization is explicitly specified (using a pragma clang loop directive)
but fails.
Avoid referencing the vtable when calling the ctor without emitting it
This fixes compilation errors about incomplete types used with WebKit's
RefPtr template. Simply calling an out of line constructor should not
instantiate all inline and defaulted virtual methods.
Tested by building and testing several big piles of code on Linux.
Summary:
This implements the -arch flag for both x86 and x86-64 by letting
them affect the default target features we pass to cc1. -m machine
flags will override the features set by -arch.
Richard Smith [Tue, 15 Jul 2014 03:37:06 +0000 (03:37 +0000)]
[modules] Rearrange and unify the way we determine if we need to pull in
redeclaration chains when pulling in a declaration. We need the redecl chain
unless we know some other declaration will trigger it to be pulled in; that
happens if our originally-canonical declaration had all the knowledge that
we have (and isn't us).
David Majnemer [Mon, 14 Jul 2014 23:40:24 +0000 (23:40 +0000)]
AST: Fix __uuidof for template specializations
While we previously supported __uuidof applied to a template
specialization, we would only find the uuid attribute if it was applied
to the template argument. We would erroneously ignore the uuid
attribute on the specialization itself.
This is required to parse Windows Runtime C++ Template Library headers.
Continue parsing an expression list even after an error is encountered.
Otherwise, multiple errors such as having unknown identifiers for two
arguments won't be diagnosed properly (e.g. only the first one would
have a diagnostic message if typo correction fails even though both
would be diagnosed if typo correction suggests a replacement).
Ben Langmuir [Mon, 14 Jul 2014 21:14:28 +0000 (21:14 +0000)]
Try harder to supress the test from r212975 on Windows
The attempt in r212980 was broken because we might not fail if
LLVM_ON_UNIX is enabled for cross compiling to Windows, and it didn't
consider mingw either.
Introduce getCorrespondingUnsignedType() in TargetInfo to work like the
corresponding AST context function, only restricted to basic integer
types. Use this to ensure getUIntPtrType() gives types consistent with
getIntPtrType(). Fix NVPTX backend to give signed intptr_t.
Richard Smith [Mon, 14 Jul 2014 19:54:05 +0000 (19:54 +0000)]
In C++98, if an rvalue reference binds to a function lvalue (or an xvalue or an
array prvalue), treat that as a direct binding. Only the class prvalue case
needs to be excluded here; the rest are extensions anyway, so we can treat them
as we would in C++11.
Alp Toker [Mon, 14 Jul 2014 18:11:38 +0000 (18:11 +0000)]
Change the diagnostic group for unsupported gcc optimizations added r212805
Use -Winvalid-command-line-argument here to align with existing gcc opt
diagnostics. Meanwhile -Wunused-command-line-argument is for flags that we
support but were, say, fed into the wrong invocation.
Richard Smith [Mon, 14 Jul 2014 02:28:44 +0000 (02:28 +0000)]
Superficial fix for PR20218: binding a function lvalue to a const reference to
a function pointer is neither better nor worse than binding a function lvalue
to a function rvalue reference. Don't get confused and think that both bindings
are binding to a function lvalue (which would make the lvalue form win); the
const reference is binding to an rvalue.
The "real" bug in PR20218 is still present: we're getting the wrong answer from
template argument deduction, and that's what leads us to this weird overload
set.
David Majnemer [Sun, 13 Jul 2014 05:19:56 +0000 (05:19 +0000)]
MS ABI: Stick internal vftables in a comdat if they have RTTI data
Previously, we would have a private backing variable and an internal
alias pointing at it.
However, -fdata-sections only fires if a global variable has non-private
linkage. This means that an unreferenced vftable wouldn't get
discarded, bloating the object file.
Instead, stick the backing variable in a comdat even if the alias has
internal linkage. This will allow the linker to drop the vftable if it
is unused.
This adds the ARM ACLE hint intrinsic wrappers to arm_acle.h. These need to be
protected with a !defined(_MSC_VER) since MSVC (and thus clang in compatibility
mode) provide these wrappers as proper builtin intrinsics.
[ASan] Collect unmangled names of global variables in Clang to print them in error reports.
Currently ASan instrumentation pass creates a string with global name
for each instrumented global (to include global names in the error report). Global
name is already mangled at this point, and we may not be able to demangle it
at runtime (e.g. there is no __cxa_demangle on Android).
Instead, create a string with fully qualified global name in Clang, and pass it
to ASan instrumentation pass in llvm.asan.globals metadata. If there is no metadata
for some global, ASan will use the original algorithm.
This fixes https://code.google.com/p/address-sanitizer/issues/detail?id=264.
clang-cl: Make all x86 CPU feature flags available, such as -msse3
Ideally, we would use the /arch cl.exe flag for this stuff.
Unfortunately, MSVC supports only 5 /arch flag values, which isn't
nearly enough to cover all the CPU features that LLVM cares about. At
the very least, we need to know about SSE3 and SSE4.1 in addition to
SSE, SSE2, AVX, and AVX2.
In the future we should add the relevant /arch mappings in addition to
these gcc-style -m flags.
MSVC accepts __noop without any trailing parens and treats it like a
literal zero. We don't treat __noop as an integer literal, but now at
least we can parse a naked __noop expression.
David Majnemer [Fri, 11 Jul 2014 20:28:10 +0000 (20:28 +0000)]
CodeGen: Don't emit a thread-wrapper if we can't touch the backing variable
OS X TLS has all accesses going through the thread-wrapper function and
gives the backing thread-local variable internal linkage. This means
that thread-wrappers must have WeakAnyLinkage so that references to the
internal thread-local variables do not get propagated to other code.
It also means that translation units which do not provide a definition
for the thread-local variable cannot attempt to emit a thread-wrapper
because the thread wrapper will attempt to reference the backing
variable.
Hal Finkel [Fri, 11 Jul 2014 17:35:21 +0000 (17:35 +0000)]
Add nonnull in CodeGen for __attribute__((nonnull))
We now have an LLVM-level nonnull attribute that can be applied to function
parameters, and we emit it for reference types (as of r209723), but did not
emit it when an __attribute__((nonnull)) was provided. Now we will.