platform/upstream/llvm.git
10 years agoConvert a vselect into a concat_vector if possible
Filipe Cabecinhas [Fri, 30 May 2014 23:03:11 +0000 (23:03 +0000)]
Convert a vselect into a concat_vector if possible

Summary:
If both vector args to vselect are concat_vectors and the condition is
constant and picks half a vector from each argument, convert the vselect
into a concat_vectors.

Added a test.

The ConvertSelectToConcatVector is assuming it doesn't get vselects with
arguments of, for example, <undef, undef, true, true>. Those get taken
care of in the checks above its call.

Reviewers: nadav, delena, grosbach, hfinkel

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D3916

llvm-svn: 209929

10 years ago[mach-o] Add support for custom sections
Nick Kledzik [Fri, 30 May 2014 22:51:04 +0000 (22:51 +0000)]
[mach-o] Add support for custom sections

llvm-svn: 209928

10 years agoisSVR4ABI() returned !isDarwin() so just move that to the else
Eric Christopher [Fri, 30 May 2014 22:47:53 +0000 (22:47 +0000)]
isSVR4ABI() returned !isDarwin() so just move that to the else
block and remove the unreachable code.

llvm-svn: 209927

10 years agoRename CreateTLOF->createTLOF to match the rest of the file and the
Eric Christopher [Fri, 30 May 2014 22:47:48 +0000 (22:47 +0000)]
Rename CreateTLOF->createTLOF to match the rest of the file and the
rest of the targets with a similar function name.

llvm-svn: 209926

10 years agoDisable stack-overflow test on ARM
Greg Fitzgerald [Fri, 30 May 2014 22:44:35 +0000 (22:44 +0000)]
Disable stack-overflow test on ARM

llvm-svn: 209925

10 years agoTake PrintingPolicy::SuppressUnwrittenScope into account when printing the
Richard Smith [Fri, 30 May 2014 22:16:51 +0000 (22:16 +0000)]
Take PrintingPolicy::SuppressUnwrittenScope into account when printing the
qualified name of a NamedDecl. Patch by Volodymyr Sapsai!

llvm-svn: 209924

10 years agoSeparate the check for blend shuffle_vector masks
Filipe Cabecinhas [Fri, 30 May 2014 21:31:21 +0000 (21:31 +0000)]
Separate the check for blend shuffle_vector masks

Summary:
Separate the check for blend shuffle_vector masks into isBlendMask.
This function will also be used to check if a vector shuffle is legal. No
change in functionality was intended, but we ended up improving codegen on
two tests, which were being (more) optimized only if the resulting shuffle
was legal.

Reviewers: nadav, delena, andreadb

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D3964

llvm-svn: 209923

10 years agoInvalidate the file system cache entries for files that may rebuild
Ben Langmuir [Fri, 30 May 2014 21:20:54 +0000 (21:20 +0000)]
Invalidate the file system cache entries for files that may rebuild

This reapplies r209910 with a fix for the assertion failures hit on the
buildbots.

original commit message:
I thought we could get away without this, but it means that the
FileEntry objects actually refer to the wrong files, since pcms are not
updated inplace, they are atomically renamed into place after compiling
a module.

So we are close to the original behaviour of invalidating the cache for
all modules being removed, but now we should only invalidate the ones
that depend on whichever module failed to load.

Unfortunately I haven't come up with a new test that didn't require
a race between parallel invocations of clang.

<rdar://problem/17038180>

llvm-svn: 209922

10 years agofixed more typos
Sanjay Patel [Fri, 30 May 2014 21:18:13 +0000 (21:18 +0000)]
fixed more typos

llvm-svn: 209921

10 years agoRoll DbgVariable::setMInsn into the constructor. No functional changes.
Adrian Prantl [Fri, 30 May 2014 21:10:13 +0000 (21:10 +0000)]
Roll DbgVariable::setMInsn into the constructor. No functional changes.

llvm-svn: 209920

10 years agoRemove unused struct from test
Alp Toker [Fri, 30 May 2014 21:07:40 +0000 (21:07 +0000)]
Remove unused struct from test

This presumably got inadvertently copied from dllimport-members.cpp.

llvm-svn: 209919

10 years agoadded link to CMake page
Sanjay Patel [Fri, 30 May 2014 21:07:25 +0000 (21:07 +0000)]
added link to CMake page

llvm-svn: 209918

10 years agoDon't use libc's "char *basename(char *)" or "char *dirname(char *)" as they are...
Greg Clayton [Fri, 30 May 2014 21:06:57 +0000 (21:06 +0000)]
Don't use libc's "char *basename(char *)" or "char *dirname(char *)" as they are not thread safe.

I switched the lldb_private::FileSpec code over to use "llvm::StringRef llvm::sys::path::filename(llvm::StringRef)" for basename() and "llvm::StringRef llvm::sys::path::parent_path(llvm::StringRef)" for dirname().

<rdar://problem/16870083>

llvm-svn: 209917

10 years agofixed typo
Sanjay Patel [Fri, 30 May 2014 20:55:55 +0000 (20:55 +0000)]
fixed typo

llvm-svn: 209916

10 years agoAdd a FIXME for a bug Doug pointed out in code review.
Richard Smith [Fri, 30 May 2014 19:44:17 +0000 (19:44 +0000)]
Add a FIXME for a bug Doug pointed out in code review.

llvm-svn: 209915

10 years agoMake bitcast, extractelement, and insertelement considered cheap for speculation.
Matt Arsenault [Fri, 30 May 2014 18:34:43 +0000 (18:34 +0000)]
Make bitcast, extractelement, and insertelement considered cheap for speculation.

This helps more branches into selects. On R600,
vectors are cheap and anything that helps
remove branches is very good.

llvm-svn: 209914

10 years agoThread Safety Analysis: fix handling of string and char literals.
DeLesley Hutchins [Fri, 30 May 2014 18:20:04 +0000 (18:20 +0000)]
Thread Safety Analysis: fix handling of string and char literals.

llvm-svn: 209913

10 years agogdb-remote signal delivery test cleanup.
Todd Fiala [Fri, 30 May 2014 17:59:47 +0000 (17:59 +0000)]
gdb-remote signal delivery test cleanup.

Learned that MacOSX only accepts signal delivery on a thread that is
already signal handling.  Reworked the test exe to cause a SIGSEGV
and recover if either nothing intercepts the SIGSEGV handler, or
if a SIGUSR1 is inserted.  The test uses the latter part to test
signal delivery on continue using the SIGUSR1.

I still don't have this working on MacOSX.  I'm seeing the
signal get delivered to a different thread than the one I'm
specifying with $Hc{thread-id} + $C{signo}, or with
$vCont;C{signo}:{thread-id};c.  I'll come back to this
after getting it working on the llgs branch on Linux x86_64.

llvm-svn: 209912

10 years agoRevert "Invalidate the file system cache entries for files that may rebuild"
Ben Langmuir [Fri, 30 May 2014 17:54:59 +0000 (17:54 +0000)]
Revert "Invalidate the file system cache entries for files that may rebuild"

This reverts commit r209910, which is breaking some of the bots.

llvm-svn: 209911

10 years agoInvalidate the file system cache entries for files that may rebuild
Ben Langmuir [Fri, 30 May 2014 17:42:15 +0000 (17:42 +0000)]
Invalidate the file system cache entries for files that may rebuild

I thought we could get away without this, but it means that the
FileEntry objects actually refer to the wrong files, since pcms are not
updated inplace, they are atomically renamed into place after compiling
a module.

So we are close to the original behaviour of invalidating the cache for
all modules being removed, but now we should only invalidate the ones
that depend on whichever module failed to load.

Unfortunately I haven't come up with a new test that didn't require
a race between parallel invocations of clang.

<rdar://problem/17038180>

llvm-svn: 209910

10 years agoLet libc++abi compile with gcc.
Nico Weber [Fri, 30 May 2014 17:27:21 +0000 (17:27 +0000)]
Let libc++abi compile with gcc.

There was a single problem in cxa_demangle.cpp, where gcc would complain
`error: changes meaning of 'String'` about the line `typedef String String;`.
According to 3.3.7p2, this diagnostic is allowed (but not required, so clang
does not have to report this).

As a fix, make string_pair a template and pass String as template parameter.
This fixes the error with gcc and also removes some repetition from the code.

No behavior change.

llvm-svn: 209909

10 years agoStart adding support for dllimport/dllexport on classes (PR11170)
Hans Wennborg [Fri, 30 May 2014 16:59:42 +0000 (16:59 +0000)]
Start adding support for dllimport/dllexport on classes (PR11170)

This implements the central part of support for dllimport/dllexport on
classes: allowing the attribute on class declarations, inheriting it
to class members, and forcing emission of exported members. It's based
on Nico Rieck's patch from http://reviews.llvm.org/D1099.

This patch doesn't propagate dllexport to bases that are template
specializations, which is an interesting problem. It also doesn't
look at the rules when redeclaring classes with different attributes,
I'd like to do that separately.

Differential Revision: http://reviews.llvm.org/D3877

llvm-svn: 209908

10 years agoFix MIPS exception personality encoding.
Logan Chien [Fri, 30 May 2014 16:48:56 +0000 (16:48 +0000)]
Fix MIPS exception personality encoding.

For MIPS, we have to encode the personality routine with
an indirect pointer to absptr; otherwise, some link warning
warning will be raised, and the program might crash in some
early MIPS Android device.

llvm-svn: 209907

10 years agoObjective-C. Diagnose assigning a block pointer type to
Fariborz Jahanian [Fri, 30 May 2014 16:35:53 +0000 (16:35 +0000)]
Objective-C. Diagnose assigning a block pointer type to
an Objective-C object type other than 'id'.
// rdar://16739120

llvm-svn: 209906

10 years agoMachO: move test to ARM directory.
Tim Northover [Fri, 30 May 2014 16:01:29 +0000 (16:01 +0000)]
MachO: move test to ARM directory.

This test specifies an ARM triple, so it needs ARM as a registered
target.

llvm-svn: 209905

10 years agoThread Safety Analysis: fix uninitialized member in TIL Literal.
DeLesley Hutchins [Fri, 30 May 2014 15:55:25 +0000 (15:55 +0000)]
Thread Safety Analysis: fix uninitialized member in TIL Literal.

llvm-svn: 209904

10 years agoPR19753: Optimize comparisons with "ashr exact" of a constanst.
Rafael Espindola [Fri, 30 May 2014 15:54:32 +0000 (15:54 +0000)]
PR19753: Optimize comparisons with "ashr exact" of a constanst.

Patch by suyog sarda.

llvm-svn: 209903

10 years ago[pr19636] Fix known bit computation in urem instruction with power of two.
Rafael Espindola [Fri, 30 May 2014 15:00:45 +0000 (15:00 +0000)]
[pr19636]  Fix known bit computation in urem instruction with power of two.
Patch by Andrey Kuharev.

llvm-svn: 209902

10 years agoSelectionDAG: skip barriers for unordered atomic operations
Tim Northover [Fri, 30 May 2014 14:41:51 +0000 (14:41 +0000)]
SelectionDAG: skip barriers for unordered atomic operations

Unordered is strictly weaker than monotonic, so if the latter doesn't have any
barriers then the former certainly shouldn't.

rdar://problem/16548260

llvm-svn: 209901

10 years agotsan: add a test from data-race-test suite:
Dmitry Vyukov [Fri, 30 May 2014 14:27:31 +0000 (14:27 +0000)]
tsan: add a test from data-race-test suite:
https://code.google.com/p/data-race-test/source/browse/trunk/unittest/racecheck_unittest.cc

llvm-svn: 209900

10 years agoAArch64: combine arm64 & aarch64 cases in switch
Tim Northover [Fri, 30 May 2014 14:14:07 +0000 (14:14 +0000)]
AArch64: combine arm64 & aarch64 cases in switch

There shouldn't be any difference in behaviour here, at least not in
any configurations people care about and possibly not in any reachable
configurations.

llvm-svn: 209899

10 years agotsan: make positive tests more robust
Dmitry Vyukov [Fri, 30 May 2014 14:08:51 +0000 (14:08 +0000)]
tsan: make positive tests more robust
Add a script that is used to deflake inherently flaky tsan tests.
It is invoked from lit tests as:
%deflake %run %t
The script runs the target program up to 10 times,
until it produces a tsan warning.

llvm-svn: 209898

10 years agotsan: optimize memory access functions
Dmitry Vyukov [Fri, 30 May 2014 13:36:29 +0000 (13:36 +0000)]
tsan: optimize memory access functions
The optimization is two-fold:
First, the algorithm now uses SSE instructions to
handle all 4 shadow slots at once. This makes processing
faster.
Second, if shadow contains the same access, we do not
store the event into trace. This increases effective
trace size, that is, tsan can remember up to 10x more
previous memory accesses.

Perofrmance impact:
Before:
[       OK ] DISABLED_BENCH.Mop8Read (2461 ms)
[       OK ] DISABLED_BENCH.Mop8Write (1836 ms)
After:
[       OK ] DISABLED_BENCH.Mop8Read (1204 ms)
[       OK ] DISABLED_BENCH.Mop8Write (976 ms)
But this measures only fast-path.
On large real applications the speedup is ~20%.

Trace size impact:
On app1:
Memory accesses                   :       1163265870
  Including same                  :        791312905 (68%)
on app2:
Memory accesses                   :        166875345
  Including same                  :        150449689 (90%)
90% of filtered events means that trace size is effectively 10x larger.

llvm-svn: 209897

10 years agoRevert "Remove redundant check discovered in post-commit review of r209505."
Daniel Jasper [Fri, 30 May 2014 13:31:14 +0000 (13:31 +0000)]
Revert "Remove redundant check discovered in post-commit review of r209505."

This breaks with MSVC.
With IsLateTemplateParsed, FunctionDecl::doesThisDeclarationHaveABody() returns true regardless of Body.

This reinstates what was fixed in r208985.

llvm-svn: 209896

10 years agoARM: use AAPCS-style prologues for embedded MachO.
Tim Northover [Fri, 30 May 2014 13:23:06 +0000 (13:23 +0000)]
ARM: use AAPCS-style prologues for embedded MachO.

Darwin prologues save their GPRs in two stages: a narrow push of r0-r7 & lr,
followed by a wide push of the remaining registers if there are any. AAPCS uses
a single push.w instruction.

It turns out that, on average, enough registers get pushed that code is smaller
in the AAPCS prologue, which is a nice property for M-class programmers. They
also have other options available for back-traces, so can hopefully deal with
the fact that FP & LR aren't adjacent in memory.

rdar://problem/15909583

llvm-svn: 209895

10 years agoMachO: support N_INDR aliases in assembly files.
Tim Northover [Fri, 30 May 2014 13:22:59 +0000 (13:22 +0000)]
MachO: support N_INDR aliases in assembly files.

This makes LLVM create N_INDR aliases (to be resolved by the linker) when
appropriate.

rdar://problem/15125513

llvm-svn: 209894

10 years agoRemove unnecessary test.
Diego Novillo [Fri, 30 May 2014 13:19:28 +0000 (13:19 +0000)]
Remove unnecessary test.

llvm-svn: 209893

10 years agoThis test is no longer relevant. It is covered already by [basic.scope.pdecl]p10...
Aaron Ballman [Fri, 30 May 2014 13:17:17 +0000 (13:17 +0000)]
This test is no longer relevant. It is covered already by [basic.scope.pdecl]p10, and our current test for that is more thorough (though our test is named p9.cpp).

llvm-svn: 209892

10 years agoThis test should no longer be XFAILed; the standard has stabilized, and the test...
Aaron Ballman [Fri, 30 May 2014 13:09:38 +0000 (13:09 +0000)]
This test should no longer be XFAILed; the standard has stabilized, and the test contents are acceptable. No diagnostics expected from this test.

llvm-svn: 209891

10 years agoSetting a heading for the section/allocate attribute so that the attribute documentat...
Aaron Ballman [Fri, 30 May 2014 12:57:13 +0000 (12:57 +0000)]
Setting a heading for the section/allocate attribute so that the attribute documentation can be properly generated.

llvm-svn: 209890

10 years ago[Sanitizers Win] Move duplicate Windows-specific compiler flags to a common CMake...
Timur Iskhodzhanov [Fri, 30 May 2014 12:42:57 +0000 (12:42 +0000)]
[Sanitizers Win] Move duplicate Windows-specific compiler flags to a common CMake variable

Reviewed at http://reviews.llvm.org/D3952

llvm-svn: 209889

10 years agoAdd a _LIBCPP_CONSTEXPR that was missed in r170026.
Nico Weber [Fri, 30 May 2014 12:09:47 +0000 (12:09 +0000)]
Add a _LIBCPP_CONSTEXPR that was missed in r170026.

(clang doesn't complain about this, but gcc does.  This is necessary for a
follow-up patch that will enable _LIBCPP_CONSTEXPR for gcc.)

llvm-svn: 209888

10 years ago[ASan Win] Fix a couple of warnings in tests and enable -WX
Timur Iskhodzhanov [Fri, 30 May 2014 11:58:32 +0000 (11:58 +0000)]
[ASan Win] Fix a couple of warnings in tests and enable -WX

llvm-svn: 209887

10 years agoImplement __divtf3 for IEEE quad precision.
Joerg Sonnenberger [Fri, 30 May 2014 11:08:18 +0000 (11:08 +0000)]
Implement __divtf3 for IEEE quad precision.

Patch by: GuanHong Liu
Differential Revision: http://reviews.llvm.org/D2800

llvm-svn: 209886

10 years agoAArch64 & ARM: disable generic test that relies on no CFG changes.
Tim Northover [Fri, 30 May 2014 10:56:12 +0000 (10:56 +0000)]
AArch64 & ARM: disable generic test that relies on no CFG changes.

llvm-svn: 209885

10 years ago[sanitizer] Fix build on ARM.
Evgeniy Stepanov [Fri, 30 May 2014 10:50:17 +0000 (10:50 +0000)]
[sanitizer] Fix build on ARM.

llvm-svn: 209884

10 years agoARM & AArch64: make use of common cmpxchg idioms after expansion
Tim Northover [Fri, 30 May 2014 10:09:59 +0000 (10:09 +0000)]
ARM & AArch64: make use of common cmpxchg idioms after expansion

The C and C++ semantics for compare_exchange require it to return a bool
indicating success. This gets mapped to LLVM IR which follows each cmpxchg with
an icmp of the value loaded against the desired value.

When lowered to ldxr/stxr loops, this extra comparison is redundant: its
results are implicit in the control-flow of the function.

This commit makes two changes: it replaces that icmp with appropriate PHI
nodes, and then makes sure earlyCSE is called after expansion to actually make
use of the opportunities revealed.

I've also added -{arm,aarch64}-enable-atomic-tidy options, so that
existing fragile tests aren't perturbed too much by the change. Many
of them either rely on undef/unreachable too pervasively to be
restored to something well-defined (particularly while making sure
they test the same obscure assert from many years ago), or depend on a
particular CFG shape, which is disrupted by SimplifyCFG.

rdar://problem/16227836

llvm-svn: 209883

10 years agoRemove redundant check discovered in post-commit review of r209505.
Daniel Jasper [Fri, 30 May 2014 09:21:17 +0000 (09:21 +0000)]
Remove redundant check discovered in post-commit review of r209505.

llvm-svn: 209882

10 years ago[ASan Win] DLL thunk: make each INTERFACE_FUNCTION unique to prevent ICF linker optim...
Timur Iskhodzhanov [Fri, 30 May 2014 09:01:17 +0000 (09:01 +0000)]
[ASan Win] DLL thunk: make each INTERFACE_FUNCTION unique to prevent ICF linker optimizations

llvm-svn: 209881

10 years agoAArch64 & ARM: remove undefined behaviour from some tests.
Tim Northover [Fri, 30 May 2014 08:59:55 +0000 (08:59 +0000)]
AArch64 & ARM: remove undefined behaviour from some tests.

llvm-svn: 209880

10 years ago[asan] Enable ASan on PowerPC.
Evgeniy Stepanov [Fri, 30 May 2014 08:52:03 +0000 (08:52 +0000)]
[asan] Enable ASan on PowerPC.

Patch by Peter Bergner.

llvm-svn: 209879

10 years agoclang/test/Frontend/Weverything-and-remarks.cpp: Add explicit -target.
NAKAMURA Takumi [Fri, 30 May 2014 07:03:16 +0000 (07:03 +0000)]
clang/test/Frontend/Weverything-and-remarks.cpp: Add explicit -target.

You could see unexpected result (*.tmp.err) with "-target x86_64-win32".

  Weverything-and-remarks.cpp:11:9: warning: 'char32_t' type specifier is incompatible with C++98 [-Wc++98-compat]
  typedef __char32_t char32_t;
          ^
  Weverything-and-remarks.cpp:12:27: error: typedef redefinition with different types ('unsigned long' vs 'unsigned long long')
  typedef long unsigned int size_t;
                            ^
  Weverything-and-remarks.cpp:19:62: warning: dynamic exception specifications are deprecated [-Wdeprecated]
    static void assign(char_type& __c1, const char_type& __c2) throw() {
                                                               ^~~~~~~
  Weverything-and-remarks.cpp:19:62: note: use 'noexcept' instead
    static void assign(char_type& __c1, const char_type& __c2) throw() {
                                                               ^~~~~~~
                                                               noexcept
  Weverything-and-remarks.cpp:25:46: warning: unused parameter '__n' [-Wunused-parameter]
                                        size_t __n) {
                                               ^
  3 warnings and 1 error generated.

llvm-svn: 209878

10 years agoTest cases named with dates is a legacy rule not used now. Rename several test cases.
Hao Liu [Fri, 30 May 2014 05:58:19 +0000 (05:58 +0000)]
Test cases named with dates is a legacy rule not used now. Rename several test cases.

llvm-svn: 209877

10 years ago[OPENMP] Improve debug location codegen for OpenMP runtime library.
Alexey Bataev [Fri, 30 May 2014 05:48:40 +0000 (05:48 +0000)]
[OPENMP] Improve debug location codegen for OpenMP runtime library.

llvm-svn: 209876

10 years ago[ELF] Fix incorrect sorting of .init_array / .fini_array sections.
Simon Atanasyan [Fri, 30 May 2014 05:29:46 +0000 (05:29 +0000)]
[ELF] Fix incorrect sorting of .init_array / .fini_array sections.

The main problem is in the predicate passed to the `std::stable_sort()`.
This predicate always returns false if **both** section's names do not
start with `.init_array` or `.fini_array` prefixes. In short, it does not
define a strict weak orderng. Suppose we have the following sections:

  .A .init_array.1 .init_array.2

The predicate states that:

  not .init_array.1 < .A
  not .A < .init_array.2
  but .init_array.1 < .init_array.2 !!!

The second problem is that `.init_array` section without number should
go last in the list. Not it has the lowest priority '0' and goes first.

The patch fixes both of the problems.

llvm-svn: 209875

10 years agoFix 'this' capturing Generic lambdas used within default initializers (PR19876)
Faisal Vali [Fri, 30 May 2014 04:39:37 +0000 (04:39 +0000)]
Fix 'this' capturing Generic lambdas used within default initializers (PR19876)

http://llvm.org/bugs/show_bug.cgi?id=19876

The following C++1y code results in a crash:

struct X {
  int m = 10;
  int n = [this](auto) { return m; }(20);
};

When implicitly instantiating the generic lambda's call operator specialization body, Sema is unable to determine the current 'this' type when transforming the MemberExpr 'm' - since it looks for the nearest enclosing FunctionDeclDC - which is obviously null.

I considered two ways to fix this:

    1) In InstantiateFunctionDefinition, when the context is saved after the lambda scope info is created, retain the 'this' pointer.
    2) Teach getCurrentThisType() to recognize it is within a generic lambda within an NSDMI/default-initializer and return the appropriate this type.

I chose to implement #2 (though I confess I do not have a compelling reason for choosing it over #1).

Richard Smith accepted the patch:
http://reviews.llvm.org/D3935

Thank you!

llvm-svn: 209874

10 years agoAllow vectorization of intrinsics such as powi,cttz and ctlz in Loop and SLP Vectorizer.
Karthik Bhat [Fri, 30 May 2014 04:31:24 +0000 (04:31 +0000)]
Allow vectorization of intrinsics such as powi,cttz and ctlz in Loop and SLP Vectorizer.

This patch adds support to vectorize intrinsics such as powi, cttz and ctlz in Vectorizer. These intrinsics are different from other
intrinsics as second argument to these function must be same in order to vectorize them and it should be represented as a scalar.
Review: http://reviews.llvm.org/D3851#inline-32769 and http://reviews.llvm.org/D3937#inline-32857

llvm-svn: 209873

10 years agoAdd a trivial ArrayRef helper overload to the SelectionDAG interface.
Chandler Carruth [Fri, 30 May 2014 03:50:06 +0000 (03:50 +0000)]
Add a trivial ArrayRef helper overload to the SelectionDAG interface.
I'm using this pretty frequently in a patch I'm working on and it seems
generally useful.

llvm-svn: 209872

10 years agoRemove use of comma operator.
Richard Trieu [Fri, 30 May 2014 03:15:17 +0000 (03:15 +0000)]
Remove use of comma operator.

llvm-svn: 209871

10 years agoWhen analyzing params/args for readnone/readonly, don't forget to consider that a...
Nick Lewycky [Fri, 30 May 2014 02:31:27 +0000 (02:31 +0000)]
When analyzing params/args for readnone/readonly, don't forget to consider that a pointer argument may be passed through a callsite to the return, and that we may need to analyze it. Fixes a bug reported on llvm-dev: lists.cs.uiuc.edu/pipermail/llvmdev/2014-May/073098.html

llvm-svn: 209870

10 years agoPR11410 - Confusing diagnostic when trailing array element tries to call deleted...
Nikola Smiljanic [Fri, 30 May 2014 01:28:28 +0000 (01:28 +0000)]
PR11410 - Confusing diagnostic when trailing array element tries to call deleted default constructor

llvm-svn: 209869

10 years ago[mach-o] Wire up mach-o binary reader to reader registry
Nick Kledzik [Fri, 30 May 2014 01:13:49 +0000 (01:13 +0000)]
[mach-o] Wire up mach-o binary reader to reader registry

llvm-svn: 209868

10 years agoPR12214 - Warn on suspicious self-compound-assignments.
Nikola Smiljanic [Fri, 30 May 2014 00:15:04 +0000 (00:15 +0000)]
PR12214 - Warn on suspicious self-compound-assignments.

llvm-svn: 209867

10 years ago[X86] Move test from r209863 to CodeGen/X86
Adam Nemet [Thu, 29 May 2014 23:52:53 +0000 (23:52 +0000)]
[X86] Move test from r209863 to CodeGen/X86

We should only run this if X86 is in the targets.

llvm-svn: 209866

10 years ago[mach-0] Add support for parsing compact unwind info section
Nick Kledzik [Thu, 29 May 2014 23:50:48 +0000 (23:50 +0000)]
[mach-0] Add support for parsing compact unwind info section

llvm-svn: 209865

10 years ago[X86] Remove AVX1 vbroadcast intrinsics
Adam Nemet [Thu, 29 May 2014 23:35:36 +0000 (23:35 +0000)]
[X86] Remove AVX1 vbroadcast intrinsics

The corresponding CFE patch replaces these intrinsics with vector initializers
in avxintrin.h.  This patch removes the LLVM intrinsics from the backend.

We now stop lowering at X86ISD::VBROADCAST custom node rather than lowering
that further to the intrinsics.

The patch only changes VBROADCASTS* and leaves VBROADCAST[FI]128 to continue
to use intrinsics.  As explained in the CFE patch, the reason is that we
currently don't generate as good code for them without the intrinsics.

CodeGen/X86/avx-vbroadcast.ll already provides coverage for this change.  It
checks that for a series of insertelements we generate the appropriate
vbroadcast instruction.

Also verified that there was no assembly change in the test-suite before and
after this patch.

llvm-svn: 209864

10 years ago[X86] Auto-upgrade AVX1 vbroadcast intrinsics
Adam Nemet [Thu, 29 May 2014 23:35:33 +0000 (23:35 +0000)]
[X86] Auto-upgrade AVX1 vbroadcast intrinsics

They are replaced with the same IR that is generated for the
vector-initializers in avxintrin.h.

The test verifies that we get back the original instruction.  I haven't seen
this approach to be used in other auto-upgrade tests (i.e. llc + FileCheck)
but I think it's the most direct way to test this case.  I believe this should
work because llc upgrades calls during parsing.  (Other tests mostly check
that assembling and disassembling yields the upgraded IR.)

llvm-svn: 209863

10 years agoMark currently failing ARM tests with XFAIL.
Greg Fitzgerald [Thu, 29 May 2014 23:34:47 +0000 (23:34 +0000)]
Mark currently failing ARM tests with XFAIL.

Differential Revision: http://reviews.llvm.org/D3857

llvm-svn: 209862

10 years agoMark unstable tests as Unsupported on ARM
Greg Fitzgerald [Thu, 29 May 2014 23:34:39 +0000 (23:34 +0000)]
Mark unstable tests as Unsupported on ARM

This is half the patch from: http://reviews.llvm.org/D3857

It lets us get deterministic results from the rest of the test suite.

llvm-svn: 209861

10 years agoAnd fix my fix to sink down through the type at the right time. My
Chandler Carruth [Thu, 29 May 2014 23:21:12 +0000 (23:21 +0000)]
And fix my fix to sink down through the type at the right time. My
original fix would actually trigger the *exact* same crasher as the
original bug for a different reason. Awesomesauce.

Working on test cases now, but wanted to get bots healthier.

llvm-svn: 209860

10 years agoPermit the "if" literal suffix with Microsoft extensions enabled.
Peter Collingbourne [Thu, 29 May 2014 23:10:15 +0000 (23:10 +0000)]
Permit the "if" literal suffix with Microsoft extensions enabled.

Differential Revision: http://reviews.llvm.org/D3963

llvm-svn: 209859

10 years ago[mach-o] Add support for parsing __eh_frame section. Generalize support for whether...
Nick Kledzik [Thu, 29 May 2014 23:07:20 +0000 (23:07 +0000)]
[mach-o] Add support for parsing __eh_frame section. Generalize support for whether symbols in a section are ignored or illegal

llvm-svn: 209858

10 years agoFix one bug in the latest incarnation of r209843 -- combining GEPs
Chandler Carruth [Thu, 29 May 2014 23:05:52 +0000 (23:05 +0000)]
Fix one bug in the latest incarnation of r209843 -- combining GEPs
across PHI nodes. The code was computing the Idxs from the 'GEP'
variable's indices when what it wanted was Op1's indices. This caused an
ASan heap-overflow for me that pin pointed the issue when Op1 had more
indices than GEP did. =] I'll let Louis add a specific test case for
this if he wants.

llvm-svn: 209857

10 years agolight up sanitizers for ARM, take 2
Greg Fitzgerald [Thu, 29 May 2014 22:38:13 +0000 (22:38 +0000)]
light up sanitizers for ARM, take 2

Differential Revision: http://reviews.llvm.org/D3794

llvm-svn: 209856

10 years agoSync list of targets with configure's reality
Jeroen Ketema [Thu, 29 May 2014 22:23:22 +0000 (22:23 +0000)]
Sync list of targets with configure's reality

llvm-svn: 209855

10 years agoLoopVectorizer: Add a check that the backedge taken count + 1 does not overflow
Arnold Schwaighofer [Thu, 29 May 2014 22:10:01 +0000 (22:10 +0000)]
LoopVectorizer: Add a check that the backedge taken count + 1 does not overflow

The loop vectorizer instantiates be-taken-count + 1 as the loop iteration count.
If this expression overflows the generated code was invalid.

In case of overflow the code now jumps to the scalar loop.

Fixes PR17288.

llvm-svn: 209854

10 years agoAdded tests for shufflevector lowering to blend instrs.
Filipe Cabecinhas [Thu, 29 May 2014 22:04:42 +0000 (22:04 +0000)]
Added tests for shufflevector lowering to blend instrs.

These tests ensure that a change I will propose in clang works as
expected.

Summary:
Added tests for the generation of blend+immediate instructions from a
shufflevector.
These tests were proposed along with a patch that was dropped. I'm
committing the tests anyway to protect against possible regressions in
codegen.

Reviewers: nadav, bkramer

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D3600

llvm-svn: 209853

10 years agoiOS simulator cleanup to make sure we use "*-apple-ios" for iOS simulator apps and...
Greg Clayton [Thu, 29 May 2014 21:33:45 +0000 (21:33 +0000)]
iOS simulator cleanup to make sure we use "*-apple-ios" for iOS simulator apps and binaries.

Changes include:
- ObjectFileMachO can now determine if a binary is "*-apple-ios" or "*-apple-macosx" by checking the min OS and SDK load commands
- ArchSpec now says "<arch>-apple-macosx" is equivalent to "<arch>-apple-ios" since the simulator mixes and matches binaries (some from the system and most from the iOS SDK).
- Getting process inforamtion on MacOSX now correctly classifies iOS simulator processes so they have "*-apple-ios" architectures in the ProcessInstanceInfo
- PlatformiOSSimulator can now list iOS simulator processes correctly instead of showing nothing by using:
    (lldb) platform select ios-simulator
    (lldb) platform process list
- debugserver can now properly return "*-apple-ios" for the triple in the process info packets for iOS simulator executables
- GDBRemoteCommunicationClient now correctly passes along the triples it gets for process info by setting the OS in the llvm::Triple correctly

<rdar://problem/17060217>

llvm-svn: 209852

10 years agoRevert "light up sanitizers for ARM"
Greg Fitzgerald [Thu, 29 May 2014 21:33:36 +0000 (21:33 +0000)]
Revert "light up sanitizers for ARM"

This commit broke the Windows build, where CMAKE_C_COMPILER can
compile and link with -march=armv7-a but the just-built-clang
cannot.

llvm-svn: 209851

10 years agoAdd more log related float constants
Jeroen Ketema [Thu, 29 May 2014 21:30:28 +0000 (21:30 +0000)]
Add more log related float constants

llvm-svn: 209850

10 years agoFix _F definitions
Jeroen Ketema [Thu, 29 May 2014 21:29:34 +0000 (21:29 +0000)]
Fix _F definitions

The 'f' was missing and, hence, the values were
considered to be doubles instead of floats.

Reviewed by: Tom Stellard

llvm-svn: 209849

10 years agoAdd definition for M_PI
Jeroen Ketema [Thu, 29 May 2014 21:24:57 +0000 (21:24 +0000)]
Add definition for M_PI

Reviewed by: Tom Stellard

llvm-svn: 209848

10 years agoThread Safety Analysis: implement review suggestions from Aaron Ballman.
DeLesley Hutchins [Thu, 29 May 2014 21:24:16 +0000 (21:24 +0000)]
Thread Safety Analysis: implement review suggestions from Aaron Ballman.

llvm-svn: 209847

10 years agoImplement AVX1 vbroadcast intrinsics with vector initializers
Adam Nemet [Thu, 29 May 2014 20:47:29 +0000 (20:47 +0000)]
Implement AVX1 vbroadcast intrinsics with vector initializers

These intrinsics are special because they directly take a memory operand (AVX2
adds the register counterparts).  Typically, other non-memop intrinsics take
registers and then it's left to isel to fold memory operands.

In order to LICM intrinsics directly reading memory, we require that no stores
are in the loop (LICM) or that the folded load accesses constant memory
(MachineLICM).  When neither is the case we fail to hoist a loop-invariant
broadcast.

We can work around this limitation if we expose the load as a regular load and
then just implement the broadcast using the vector initializer syntax.  This
exposes the load to LICM and other optimizations.

At the IR level this is translated into a series of insertelements.  The
sequence is already recognized as a broadcast so there is no impact on the
quality of codegen.

_mm256_broadcast_pd and _mm256_broadcast_ps are not updated by this patch
because right now we lack the DAG-combiner smartness to recover the broadcast
instructions.  This will be tackled in a follow-on.

There will be completing changes on the LLVM side to remove the LLVM
intrinsics and to auto-upgrade bitcode files.

Fixes <rdar://problem/16494520>

llvm-svn: 209846

10 years agogdb-remote testing: new test, cleaned up socket reading.
Todd Fiala [Thu, 29 May 2014 20:44:45 +0000 (20:44 +0000)]
gdb-remote testing: new test, cleaned up socket reading.

Added new SocketPacketPump class to decouple gdb remote packet
reading from packet expectations code.  This allowed for cleaner
implementation of the separate $O output streams (non-deterministic
packaging of inferior stdout/stderr) from all the rest of the packets.

Added a packet expectation matcher that can match expected accumulated
output with a timeout.  Use a dictionary with "type":"output_match".
See lldbgdbserverutils.MatchRemoteOutputEntry for details.

Added a gdb remote test to verify that $Hc (continue thread selection)
plus signal delivery ($C{signo}) works.  Having trouble getting this
to pass with debugserver on MacOSX 10.9.  Tried different variants,
including $vCont;C{signo}:{thread-id};c.  In some cases, I get the
test exe's signal handler to run ($vCont variant first time), in others I don't
($vCont second and further times).  $C{signo} doesn't hit the signal
handler code at all in the test exe but delivers a stop.  Further
$Hc and $C{signo} deliver the stop marking the wrong thread.  For now I'm
marking the test as XFAIL on dsym/debugserver.  Will revisit this on
lldb-dev.

Updated the text exe for these tests to support thread:print-ids (each
thread announces its thread id) and provide a SIGUSR1 thread handler
that prints out the thread id on which it was signaled.

llvm-svn: 209845

10 years ago[mach-o] Add support for parsing CFString sections
Nick Kledzik [Thu, 29 May 2014 20:44:21 +0000 (20:44 +0000)]
[mach-o] Add support for parsing CFString sections

llvm-svn: 209844

10 years agoAdd support for combining GEPs across PHI nodes
Louis Gerbarg [Thu, 29 May 2014 20:29:47 +0000 (20:29 +0000)]
Add support for combining GEPs across PHI nodes

Currently LLVM will generally merge GEPs. This allows backends to use more
complex addressing modes. In some cases this is not happening because there
is PHI inbetween the two GEPs:

  GEP1--\
        |-->PHI1-->GEP3
  GEP2--/

This patch checks to see if GEP1 and GEP2 are similiar enough that they can be
cloned (GEP12) in GEP3's BB, allowing GEP->GEP merging (GEP123):

  GEP1--\                     --\                           --\
        |-->PHI1-->GEP3  ==>    |-->PHI2->GEP12->GEP3 == >    |-->PHI2->GEP123
  GEP2--/                     --/                           --/

This also breaks certain use chains that are preventing GEP->GEP merges that the
the existing instcombine would merge otherwise.

Tests included.

llvm-svn: 209843

10 years agoThread Safety Analysis: minor changes to TIL pretty-printing.
DeLesley Hutchins [Thu, 29 May 2014 20:28:53 +0000 (20:28 +0000)]
Thread Safety Analysis: minor changes to TIL pretty-printing.

llvm-svn: 209842

10 years agoAdd documentation for -Rpass*
Diego Novillo [Thu, 29 May 2014 20:13:27 +0000 (20:13 +0000)]
Add documentation for -Rpass*

Summary:
This adds documentation for -Rpass, -Rpass-missed and -Rpass-analysis.
It also adds release notes for 3.5.

Reviewers: rsmith

Subscribers: cfe-commits

Differential Revision: http://reviews.llvm.org/D3730

llvm-svn: 209841

10 years agoUpdate Credits.
Chad Rosier [Thu, 29 May 2014 19:59:58 +0000 (19:59 +0000)]
Update Credits.

llvm-svn: 209840

10 years agoAdd flags -Rpass-missed and -Rpass-analysis.
Diego Novillo [Thu, 29 May 2014 19:55:06 +0000 (19:55 +0000)]
Add flags -Rpass-missed and -Rpass-analysis.

Summary:
These two flags are in the same family as -Rpass, but are used in
different situations.

-Rpass-missed is used by optimizers to inform the user when they tried
to apply an optimization but couldn't (or wouldn't).

-Rpass-analysis is used by optimizers to report analysis results back
to the user (e.g., why the transformation could not be applied).

Depends on D3682.

Reviewers: rsmith

Subscribers: cfe-commits

Differential Revision: http://reviews.llvm.org/D3683

llvm-svn: 209839

10 years agoimplement missing SCEVDivision case
Sebastian Pop [Thu, 29 May 2014 19:44:09 +0000 (19:44 +0000)]
implement missing SCEVDivision case

without this case we would end on an infinite recursion: the remainder is zero,
so Numerator - Remainder is equal to Numerator and so we would recursively ask
for the division of Numerator by Denominator.

llvm-svn: 209838

10 years agofail to find dimensions when ElementSize is nullptr
Sebastian Pop [Thu, 29 May 2014 19:44:05 +0000 (19:44 +0000)]
fail to find dimensions when ElementSize is nullptr

when ScalarEvolution::getElementSize returns nullptr it is safe to early return
in ScalarEvolution::findArrayDimensions such that we avoid later problems when
we try to divide the terms by ElementSize.

llvm-svn: 209837

10 years agoMake the -mno-global-merge option work for arm64/aarch64. rdar://17024719
Bob Wilson [Thu, 29 May 2014 19:43:02 +0000 (19:43 +0000)]
Make the -mno-global-merge option work for arm64/aarch64. rdar://17024719

llvm-svn: 209836

10 years agolight up sanitizers for ARM
Greg Fitzgerald [Thu, 29 May 2014 19:01:32 +0000 (19:01 +0000)]
light up sanitizers for ARM

You can expect the sanitizers to be built under any of the following conditions:

1) CMAKE_C_COMPILER is GCC built to cross-compile to ARM
2) CMAKE_C_COMPILER is Clang built to cross-compile to ARM (ARM is default target)
3) CMAKE_C_COMPILER is Clang and CMAKE_C_FLAGS contains -target and --sysroot

Differential Revision: http://reviews.llvm.org/D3794

llvm-svn: 209835

10 years agoUse range-based for loops in ASan, TSan and MSan
Alexey Samsonov [Thu, 29 May 2014 18:40:48 +0000 (18:40 +0000)]
Use range-based for loops in ASan, TSan and MSan

llvm-svn: 209834

10 years agoFixed the Module::Module(ModuleSpec) constructor to properly copy the file offset...
Greg Clayton [Thu, 29 May 2014 17:52:46 +0000 (17:52 +0000)]
Fixed the Module::Module(ModuleSpec) constructor to properly copy the file offset and object file mod time from the actual module specifications so we will always be able to directly load the image we care about when calling Module::GetObjectFile().

llvm-svn: 209833

10 years ago[ADT] Delete the Twine assignment operator
Reid Kleckner [Thu, 29 May 2014 17:12:05 +0000 (17:12 +0000)]
[ADT] Delete the Twine assignment operator

This makes it slightly harder to misuse Twines.  It is still possible to
refer to destroyed temporaries with the regular constructors, though.

Patch by Marco Alesiani!

llvm-svn: 209832

10 years agoMake runlines consistent with other dll attribute tests
Nico Rieck [Thu, 29 May 2014 16:57:34 +0000 (16:57 +0000)]
Make runlines consistent with other dll attribute tests

llvm-svn: 209831

10 years agoFix some minor typos in tests
Nico Rieck [Thu, 29 May 2014 16:57:18 +0000 (16:57 +0000)]
Fix some minor typos in tests

llvm-svn: 209830