No intended behavior change on Linux and Mac OS X. On Windows, having libcxxabi
in one's checkout should now no longer break running cmake. (cl.exe supports
c++11, but doesn't understand a -std= flag.)
llvm-svn: 213214
When plugins are used the Multiplex(AST)Consumer is employed to dispatch
to both the plugin ASTConsumers and the IRGen ASTConsumer. It wasn't
dispatching a critical call for debug info, resulting in plugin users
having a negative debugging experience.
While I'm here, forward a bunch of other missing calls through the
consumer that seem like they should be there.
To test this, use the example plugin (requires plugins and examples) and
split the test case up so that the plugin testing can be done under that
requirement while the non-plugin testing will execute even in builds
that don't include plugin support or examples.
llvm-svn: 213213
By having the two variables 'a' and 'b' in this test in a namespace, the
type was required to be complete before any debug info was ever emitted
(the entire namespace is parsed before the variables were emitted), this
meant that the codepath in which a declaration is emitted, then later on
the type is required to be complete and the debug info must be upgraded
to a definition was not used.
Moving the variables outside a namespace fixes this test coverage bug.
(interestingly, code coverage didn't help here -
HandleTagDeclRequireDefinition is fully covered because it's called even
in cases where the type hasn't been emitted for debug info at all
(further down in CGDebugInfo this no-ops) - so CC wouldn't've helped
catch this test coverage problem)
llvm-svn: 213211
Summary:
Converting outermost zext(a) to sext(a) causes worse code when the
computation of zext(a) could be reused. For example, after converting
... = array[zext(a)]
... = array[zext(a) + 1]
to
... = array[sext(a)]
... = array[zext(a) + 1],
the program computes sext(a), which is actually unnecessary. I added one
test in split-gep-and-gvn.ll to illustrate this scenario.
Also, with r211281 and r211084, we annotate more "nuw" tags to
computation involving CUDA intrinsics such as threadIdx.x. These
annotations help with splitting GEP a lot, rendering the benefit we get
from this reverted optimization only marginal.
Test Plan: make check-all
Reviewers: eliben, meheff
Reviewed By: meheff
Subscribers: jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D4542
llvm-svn: 213209
Clang supports __assume, at least at the semantic level, when MS extensions are
enabled. Unfortunately, trying to actually compile code using __assume would
result in this error:
error: cannot compile this builtin function yet
__assume is an optimizer hint, and can be ignored at the IR level. Until LLVM
supports assumptions at the IR level, a noop lowering is valid, and that is
what is done here.
llvm-svn: 213206
This fixes an issue where a local value is defined before and used after an
inline asm call with side effects.
This fix simply flushes the local value map, which updates the insertion point
for the inline asm call to be above any previously defined local values.
This fixes <rdar://problem/17694203>
llvm-svn: 213203
constructor (and pass it an implicitly-generated std::initializer_list object),
be sure to mark the resulting construction as list-initialization. This fixes
an assert in template instantiation where we previously thought we'd got direct
non-list initialization without any parentheses.
llvm-svn: 213201
Also track down the required binary by trying to locate the main executable module through LLDB's symbol and executable file locating code.
<rdar://problem/16570258>
llvm-svn: 213199
however certain sloppy Makefiles pass -z options directly to
the compiler. This patch enables clang to recognize these
options (because -z is not used by clang itself).
llvm-svn: 213198
The problem was that we have an IOHandler thread that services the IOHandler stack. The command interepter is on the top of the stack and it receives a "expression ..." command, and it calls the IOHandlerIsComplete() callback in the command interpereter delegate which runs an expression. This causes the IOHandlerProcessSTDIO to be pushed, but since we are running the code from the IOHandler thread, it won't get run. When CTRL+C is pressed, we do deliver the interrupt to the IOHandlerProcessSTDIO::Interrupt() function, but it was always writing 'i' to the interrupt pipe, even if we weren't actively reading from the debugger input and the pipes. This fix works around the issue by directly issuing the async interrupt to the process if the process is running.
A longer term more correct fix would to be run the IOHandler thread and have it just do the determination of the input and when complete input is received, run the code that handles that input on another thread and syncronize with that other thread to detect when more input is desired. That change is too big to make right now, so this fix will tide us over until we can get there.
<rdar://problem/16556228>
llvm-svn: 213196
All architecture specific handling is now done in the appropriate
ArchHandler subclass.
The StubsPass and GOTPass have been simplified. All architecture specific
variations in stubs are now encoded in a table which is vended by the
current ArchHandler.
llvm-svn: 213187
We were not considering the stated alignment on vector loads/stores,
leading us to generate vector instructions even when we do not have
sufficient alignment.
Now, for IR like:
%1 = load <4 x float>, <4 x float>* %ptr, align 4
we will generate correct, conservative PTX like:
ld.f32 ... [%ptr]
ld.f32 ... [%ptr+4]
ld.f32 ... [%ptr+8]
ld.f32 ... [%ptr+12]
Or if we have an alignment of 8 (for example), we can
generate code like:
ld.v2.f32 ... [%ptr]
ld.v2.f32 ... [%ptr+8]
llvm-svn: 213186
This value gets set to a max uint32_t value when there is no known limit; otherwise,
it is set to a value appropriate for the platform. For the moment, only
Linux, FreeBSD and NetBSD set it to 16. All other platforms set it to
the max uint32_t value.
Modifies the Process private state thread names to fit within a 16-character limit
when the max thread name length is <= 16. These guarantee that the thread names
can be distinguished within the first 16 characters. Prior to this change, those
threads had names in the final dotted name segment that were not distinguishable
within the first 16 characters.
llvm-svn: 213183
r213171 renames the 'clangRewriteCore' library to 'clangRewrite'.
This change simply updates the makefiles to reference the correct
library name.
llvm-svn: 213181
With this change the memory of buffer in NestedNameSpecifierLocBuilder
is allocated in one place. It also prevents from allocation of tiny blocks.
llvm-svn: 213178
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.
llvm-svn: 213171
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
llvm-svn: 213167
This change adds a member to the base test case for gdb-remote that
indicates whether a stub makes two X stop notification reports on kill
commands. This is set to true for debugserver tests.
The test for killing an attached process after it's first stop notification
has been modified to look at that flag and add an extra X packet matcher
so the "unmatched packet warning" doesn't get emitted for the second X on
MacOSX with debugserver.
I also broke those tests out of the monolithic TestLldbGdbServer mega test
case and put it in its own, new TestGdbRemoteKill.py file and test case.
Tested:
Ubuntu 14.04 x86_64, clang-3.5 built lldb, no test failures.
MacOSX 10.9.4, Xcode 6.0 Beta 3 built lldb, no test failures.
llvm-svn: 213166
It turns out that in most cases (the main exception being i1-related
types) once these operations are formed we cannot separate them and
the targets end up having to deal with them whether they want to or
not.
This is not a good situation, and a more reasonable default can be
formed by ackowledging this and having targets leave them as Legal.
Only x86 seems to be affected (other targets don't even try marking
the operation Expand).
Mostly there's no visible change here yet, but it will be useful to
have truly expanded EXTLOADS for MVT::f16 softening support.
llvm-svn: 213162
Nico Weber