platform/upstream/llvm.git
11 years agoUpdate the intro paragraph of text describing the table below to
Jason Molenda [Fri, 5 Oct 2012 04:28:33 +0000 (04:28 +0000)]
Update the intro paragraph of text describing the table below to
reflect that the columns were just swapped.

llvm-svn: 165295

11 years agoSwap the columns in the lldb & gdb command comparison table. I
Jason Molenda [Fri, 5 Oct 2012 04:26:06 +0000 (04:26 +0000)]
Swap the columns in the lldb & gdb command comparison table.  I
often hear feedback from people that having the lldb commands on
the left hand side is non-intuitive -- they say "I want a *gdb* to
*lldb* table, not a *lldb* to *gdb* table".  It doesn't seem odious
to go from the right-hand column to the left hand column but I've
heard this same feedback enough that it's also pointless to keep
the format lldb-first, gdb-second when it was an arbitrary choice
to begin with.

llvm-svn: 165294

11 years agodocs: Add HowToSetUpLLVMStyleRTTI.rst.
Sean Silva [Fri, 5 Oct 2012 03:32:01 +0000 (03:32 +0000)]
docs: Add HowToSetUpLLVMStyleRTTI.rst.

This document describes how to set up LLVM-style RTTI for a class
hierarchy. Surprisingly, this was not previously documented.

Also, link it into ProgrammersManual.html.

llvm-svn: 165293

11 years agotblgen: Use appropriate LLVM-style RTTI functions.
Sean Silva [Fri, 5 Oct 2012 03:32:00 +0000 (03:32 +0000)]
tblgen: Use appropriate LLVM-style RTTI functions.

Use isa<> or cast<> when semantically that is what is happening. Also
some trivial "style" cleanups at fix sites.

llvm-svn: 165292

11 years agotblgen: Replace uses of dynamic_cast<XXXRecTy> with dyn_cast<>.
Sean Silva [Fri, 5 Oct 2012 03:31:58 +0000 (03:31 +0000)]
tblgen: Replace uses of dynamic_cast<XXXRecTy> with dyn_cast<>.

This is a mechanical change of dynamic_cast<> to dyn_cast<>. A number of
these uses are actually more like isa<> or cast<>, and will be changed
to the semanticaly appropriate one in a future patch.

llvm-svn: 165291

11 years agotblgen: Put dyn_cast<> infrastructure in place for RecTy hierarchy.
Sean Silva [Fri, 5 Oct 2012 03:31:56 +0000 (03:31 +0000)]
tblgen: Put dyn_cast<> infrastructure in place for RecTy hierarchy.

llvm-svn: 165290

11 years agoMake sure to generate the right kind of MDNode for enum forward declarations.
Eli Friedman [Fri, 5 Oct 2012 01:49:33 +0000 (01:49 +0000)]
Make sure to generate the right kind of MDNode for enum forward declarations.
PR14029, clang part.

llvm-svn: 165289

11 years agoMake sure to generate the right kind of MDNode for enum forward declarations.
Eli Friedman [Fri, 5 Oct 2012 01:49:14 +0000 (01:49 +0000)]
Make sure to generate the right kind of MDNode for enum forward declarations.
PR14029, LLVM part.

llvm-svn: 165288

11 years agoFollow up to r165072. Try a different approach: only move the load when it's going...
Evan Cheng [Fri, 5 Oct 2012 01:48:22 +0000 (01:48 +0000)]
Follow up to r165072. Try a different approach: only move the load when it's going to be folded into the call. rdar://12437604

llvm-svn: 165287

11 years agoWorkaround for libstdc++4.6 <atomic> bug: make comment more explicit about what's...
Richard Smith [Fri, 5 Oct 2012 01:46:25 +0000 (01:46 +0000)]
Workaround for libstdc++4.6 <atomic> bug: make comment more explicit about what's going on, per Sean Silva's suggestion.

llvm-svn: 165286

11 years agoTeach the new SROA a new trick. Now we zap any memcpy or memmoves which
Chandler Carruth [Fri, 5 Oct 2012 01:29:09 +0000 (01:29 +0000)]
Teach the new SROA a new trick. Now we zap any memcpy or memmoves which
are in fact identity operations. We detect these and kill their
partitions so that even splitting is unaffected by them. This is
particularly important because Clang relies on emitting identity memcpy
operations for struct copies, and these fold away to constants very
often after inlining.

Fixes the last big performance FIXME I have on my plate.

llvm-svn: 165285

11 years agoLift the speculation visitor above all the helpers that are targeted at
Chandler Carruth [Fri, 5 Oct 2012 01:29:06 +0000 (01:29 +0000)]
Lift the speculation visitor above all the helpers that are targeted at
the rewrite visitor to make the fact that the speculation is completely
independent a bit more clear.

I promise that this is just a cut/paste of the one visitor and adding
the annonymous namespace wrappings. The diff may look completely
preposterous, it does in git for some reason.

llvm-svn: 165284

11 years agoImplement -Wshift-op-parentheses for: a << b + c
David Blaikie [Fri, 5 Oct 2012 00:41:03 +0000 (00:41 +0000)]
Implement -Wshift-op-parentheses for: a << b + c

This appears to be consistent with GCC's implementation of the same warning
under -Wparentheses. Suppressing a << b + c for cases where 'a' is a user
defined type for compatibility with C++ stream IO. Otherwise suggest
parentheses around the addition or subtraction subexpression.

(this came up when MSVC was complaining (incorrectly, so far as I can tell)
about a perceived violation of this within the LLVM codebase, PR14001)

llvm-svn: 165283

11 years agoUse -object_path_lto when linking executables if building Apple style.
Bill Wendling [Fri, 5 Oct 2012 00:22:46 +0000 (00:22 +0000)]
Use -object_path_lto when linking executables if building Apple style.

llvm-svn: 165282

11 years ago[libclang] Now that we have a CXModule object, pass it to the
Argyrios Kyrtzidis [Fri, 5 Oct 2012 00:22:40 +0000 (00:22 +0000)]
[libclang] Now that we have a CXModule object, pass it to the
importedASTFile indexing callback.

llvm-svn: 165281

11 years ago[libclang] Introduce new C functions that provide information about modules:
Argyrios Kyrtzidis [Fri, 5 Oct 2012 00:22:37 +0000 (00:22 +0000)]
[libclang] Introduce new C functions that provide information about modules:

clang_Cursor_getModule
clang_Module_getParent
clang_Module_getName
clang_Module_getFullName
clang_Module_getNumTopLevelHeaders
clang_Module_getTopLevelHeader

llvm-svn: 165280

11 years ago[Modules] Introduce Module::TopHeaders which is a set of top-level headers
Argyrios Kyrtzidis [Fri, 5 Oct 2012 00:22:33 +0000 (00:22 +0000)]
[Modules] Introduce Module::TopHeaders which is a set of top-level headers
that are associated with a (sub)module.

llvm-svn: 165279

11 years ago[preprocessing record] Have PPEntityID be independent of the size of the
Argyrios Kyrtzidis [Fri, 5 Oct 2012 00:22:28 +0000 (00:22 +0000)]
[preprocessing record] Have PPEntityID be independent of the size of the
loaded entities vector, otherwise its meaning will change when a module
is imported and the vector size changes.

llvm-svn: 165278

11 years ago[libclang] Introduce CXCursor_ModuleImportDecl cursor kind, used for a module
Argyrios Kyrtzidis [Fri, 5 Oct 2012 00:22:24 +0000 (00:22 +0000)]
[libclang] Introduce CXCursor_ModuleImportDecl cursor kind, used for a module
import declaration.

llvm-svn: 165277

11 years agoTest case for r165275.
Chad Rosier [Thu, 4 Oct 2012 23:59:54 +0000 (23:59 +0000)]
Test case for r165275.

llvm-svn: 165276

11 years ago[ms-inline asm] Add support for parsing [Intel dialect] memory operands that use
Chad Rosier [Thu, 4 Oct 2012 23:59:38 +0000 (23:59 +0000)]
[ms-inline asm] Add support for parsing [Intel dialect] memory operands that use
segmented registers.  Test case to come.

llvm-svn: 165275

11 years ago<rdar://problem/12099999> renaming SBStream::Printf to Print in the scripting world...
Enrico Granata [Thu, 4 Oct 2012 23:54:09 +0000 (23:54 +0000)]
<rdar://problem/12099999> renaming SBStream::Printf to Print in the scripting world in order to avoid supporting varargs through SWIG

llvm-svn: 165274

11 years agoIf we flow off the end of a value-returning function:
Richard Smith [Thu, 4 Oct 2012 23:52:29 +0000 (23:52 +0000)]
If we flow off the end of a value-returning function:
 - outside C++, return undef (behavior is not undefined unless the value is used)
 - in C++, with -fcatch-undefined-behavior, perform an appropriate trap
 - in C++, produce an 'unreachable' (behavior is undefined immediately)

llvm-svn: 165273

11 years agoA tweak to the previous commit to ensure that we don't try to use -> on a NULL pointe...
Enrico Granata [Thu, 4 Oct 2012 23:17:01 +0000 (23:17 +0000)]
A tweak to the previous commit to ensure that we don't try to use -> on a NULL pointer (should not happen but better be safe than sorry)

llvm-svn: 165272

11 years ago<rdar://problem/12413390> Fixing an issue where synthetic ValueObjects do not properl...
Enrico Granata [Thu, 4 Oct 2012 23:13:00 +0000 (23:13 +0000)]
<rdar://problem/12413390> Fixing an issue where synthetic ValueObjects do not properly resolve their value

llvm-svn: 165271

11 years agoRename the Target specific passes in the DataLayout class to be Target agnostic.
Micah Villmow [Thu, 4 Oct 2012 23:01:22 +0000 (23:01 +0000)]
Rename the Target specific passes in the DataLayout class to be Target agnostic.

llvm-svn: 165270

11 years agoRan the sources through the compiler with -Wshadow warnings
Jason Molenda [Thu, 4 Oct 2012 22:47:07 +0000 (22:47 +0000)]
Ran the sources through the compiler with -Wshadow warnings
enabled after we'd found a few bugs that were caused by shadowed
local variables; the most important issue this turned up was
a common mistake of trying to obtain a mutex lock for the scope
of a code block by doing

        Mutex::Locker(m_map_mutex);

This doesn't assign the lock object to a local variable; it is
a temporary that has its dtor called immediately.  Instead,

        Mutex::Locker locker(m_map_mutex);

does what is intended.  For some reason -Wshadow happened to
highlight these as shadowed variables.

I also fixed a few obivous and easy shadowed variable issues
across the code base but there are a couple dozen more that
should be fixed when someone has a free minute.
<rdar://problem/12437585>

llvm-svn: 165269

11 years agoRemoved a directive to delete the test subdirectories
Sean Callanan [Thu, 4 Oct 2012 22:37:53 +0000 (22:37 +0000)]
Removed a directive to delete the test subdirectories
from LLVM and Clang.  This made "svn update" very
unpleasant if the original repository was fetched by
build-llvm.pl.

llvm-svn: 165268

11 years agoWhen merging connsecutive stores, use vectors to store the constant zero.
Nadav Rotem [Thu, 4 Oct 2012 22:35:15 +0000 (22:35 +0000)]
When merging connsecutive stores, use vectors to store the constant zero.

llvm-svn: 165267

11 years agoMade the i386 ABI mark EBP as non-volatile,
Sean Callanan [Thu, 4 Oct 2012 22:24:27 +0000 (22:24 +0000)]
Made the i386 ABI mark EBP as non-volatile,
because the unwinders typically can find its
value.

llvm-svn: 165266

11 years agolibstdc++-4.6 needs the same common_type fix as libstdc++-4.7. Other than that,
Richard Smith [Thu, 4 Oct 2012 22:23:07 +0000 (22:23 +0000)]
libstdc++-4.6 needs the same common_type fix as libstdc++-4.7. Other than that,
Clang can now cope with its eccentricities in C++11 mode.

llvm-svn: 165265

11 years agoClangASTContext::GetIndexOfChildWithName - increment the child index we plan to retur...
Jim Ingham [Thu, 4 Oct 2012 22:22:16 +0000 (22:22 +0000)]
ClangASTContext::GetIndexOfChildWithName - increment the child index we plan to return as we iterate through the ivars.

<rdar://problem/12433299>

llvm-svn: 165264

11 years agoEgriegious hack to support libstdc++4.6's broken <atomic> header, which defines
Richard Smith [Thu, 4 Oct 2012 22:13:39 +0000 (22:13 +0000)]
Egriegious hack to support libstdc++4.6's broken <atomic> header, which defines
a non-inline namespace, then reopens it as inline to try to add its symbols to
the surrounding namespace. In this one special case, permit the namespace to be
reopened as inline, and patch up the name lookup tables to match.

llvm-svn: 165263

11 years agoResubmit the copying of TargetData to DataLayout without any changes to the files...
Micah Villmow [Thu, 4 Oct 2012 22:08:14 +0000 (22:08 +0000)]
Resubmit the copying of TargetData to DataLayout without any changes to the files, this should fix the problems and the changes to rename to DataLayout will come next.

llvm-svn: 165262

11 years agoFixed an assertion in the SymbolFile resulting
Sean Callanan [Thu, 4 Oct 2012 22:06:29 +0000 (22:06 +0000)]
Fixed an assertion in the SymbolFile resulting
from a NULL ObjCInterfaceDecl.

llvm-svn: 165261

11 years ago<rdar://problem/12424824> Making sure that we correctly update our synthetic children...
Enrico Granata [Thu, 4 Oct 2012 21:46:06 +0000 (21:46 +0000)]
<rdar://problem/12424824> Making sure that we correctly update our synthetic children provider for NSDictionary - providing better support for dynamic types by letting the filter recalculate itself when the type of the object changes

llvm-svn: 165260

11 years agoFixed FunctionTypeLoc source range.
Abramo Bagnara [Thu, 4 Oct 2012 21:42:10 +0000 (21:42 +0000)]
Fixed FunctionTypeLoc source range.

llvm-svn: 165259

11 years agoFixed instantiated operators source range.
Abramo Bagnara [Thu, 4 Oct 2012 21:40:42 +0000 (21:40 +0000)]
Fixed instantiated operators source range.

llvm-svn: 165258

11 years agoFixed friend decl source range.
Abramo Bagnara [Thu, 4 Oct 2012 21:39:47 +0000 (21:39 +0000)]
Fixed friend decl source range.

llvm-svn: 165257

11 years agoFixed ParamDecl source range for implicit typed k&r parameters.
Abramo Bagnara [Thu, 4 Oct 2012 21:38:29 +0000 (21:38 +0000)]
Fixed ParamDecl source range for implicit typed k&r parameters.

llvm-svn: 165256

11 years agoThis patch corrects commit 165126 by using an integer bit width instead of
Preston Gurd [Thu, 4 Oct 2012 21:33:40 +0000 (21:33 +0000)]
This patch corrects commit 165126 by using an integer bit width instead of
a pointer to a type, in order to remove the uses of getGlobalContext().

Patch by Tyler Nowicki.

llvm-svn: 165255

11 years agoARM: locate user-defined text sections next to default text.
Jim Grosbach [Thu, 4 Oct 2012 21:33:24 +0000 (21:33 +0000)]
ARM: locate user-defined text sections next to default text.

Make sure functions located in user specified text sections (via the
section attribute) are located together with the default text sections.
Otherwise, for large object files, the relocations for call instructions
are more likely to be out of range. This becomes even more likely in the
presence of LTO.

rdar://12402636

llvm-svn: 165254

11 years agoBacking out my changes, something screwed up from my patches, starting over.
Micah Villmow [Thu, 4 Oct 2012 21:08:50 +0000 (21:08 +0000)]
Backing out my changes, something screwed up from my patches, starting over.

llvm-svn: 165253

11 years ago<rdar://problem/12424824> Making sure that we correctly update our synthetic children...
Enrico Granata [Thu, 4 Oct 2012 21:04:46 +0000 (21:04 +0000)]
<rdar://problem/12424824> Making sure that we correctly update our synthetic children provider for NSArray - the same work will need to be done for NSDictionary

llvm-svn: 165252

11 years agoRename TargetData to DataLayout in DataLayout.cpp. This should fix a build failure...
Micah Villmow [Thu, 4 Oct 2012 20:49:16 +0000 (20:49 +0000)]
Rename TargetData to DataLayout in DataLayout.cpp. This should fix a build failure from r165249 where the wrong version of the file was submitted.

llvm-svn: 165251

11 years agoUpdate this a bit more to represent how the prologue should work:
Eric Christopher [Thu, 4 Oct 2012 20:46:14 +0000 (20:46 +0000)]
Update this a bit more to represent how the prologue should work:

a) frame setup instructions define the prologue
b) we shouldn't change our location mid-stream

Add a test to make sure that the stack adjustment stays within
the prologue.

llvm-svn: 165250

11 years agoCreate the DataLayout class, as a direct copy of TargetData.
Micah Villmow [Thu, 4 Oct 2012 20:44:22 +0000 (20:44 +0000)]
Create the DataLayout class, as a direct copy of TargetData.

llvm-svn: 165249

11 years agoFix the CodeGen/ppc64-varargs-struct.c test case to tolerate release builds.
Bill Schmidt [Thu, 4 Oct 2012 20:33:57 +0000 (20:33 +0000)]
Fix the CodeGen/ppc64-varargs-struct.c test case to tolerate release builds.

llvm-svn: 165247

11 years agoAdding MCJIT and MemoryBuffer unit tests
Andrew Kaylor [Thu, 4 Oct 2012 20:29:44 +0000 (20:29 +0000)]
Adding MCJIT and MemoryBuffer unit tests

Patch by Daniel Malea.

llvm-svn: 165246

11 years agoAdd a test case for r156143, which enabled general varargs support for the
Bill Schmidt [Thu, 4 Oct 2012 20:18:38 +0000 (20:18 +0000)]
Add a test case for r156143, which enabled general varargs support for the
64-bit PPC SVR4 ABI.

The test verifies passing of structures, items with 16-byte alignment, and
small items that are passed right-justified in the parameter save area slot.

llvm-svn: 165245

11 years agoAccidently checked in the files, only wanted to copy them.
Micah Villmow [Thu, 4 Oct 2012 20:08:28 +0000 (20:08 +0000)]
Accidently checked in the files, only wanted to copy them.

llvm-svn: 165244

11 years ago(no commit message)
Micah Villmow [Thu, 4 Oct 2012 20:06:07 +0000 (20:06 +0000)]
(no commit message)

llvm-svn: 165243

11 years ago(no commit message)
Micah Villmow [Thu, 4 Oct 2012 20:05:12 +0000 (20:05 +0000)]
(no commit message)

llvm-svn: 165242

11 years agoAdd register encoding support in X86 backend
Michael Liao [Thu, 4 Oct 2012 19:50:43 +0000 (19:50 +0000)]
Add register encoding support in X86 backend

- Add 'HwEncoding' for X86 registers and call getEncodingValue() to
  retrieve their encoding values.
- This's the first step to adopt new scheme. Furthur revising is onging.

llvm-svn: 165241

11 years agoDriver: Link crtfastmath.o if it's available and -ffast-math is specified.
Benjamin Kramer [Thu, 4 Oct 2012 19:42:20 +0000 (19:42 +0000)]
Driver: Link crtfastmath.o if it's available and -ffast-math is specified.

crtfastmath.o contains routines to set the floating point flags to a faster,
unsafe mode. Linking it in speeds up code dealing with denormals significantly
(PR14024).

For now this is only enabled on linux where I can test it and crtfastmath.o is
widely available. We may want to provide a similar file with compiler-rt
eventually and/or enable it on other platforms too.

llvm-svn: 165240

11 years agoFix doxygen comment to match parameters' names.
Jakub Staszak [Thu, 4 Oct 2012 19:10:44 +0000 (19:10 +0000)]
Fix doxygen comment to match parameters' names.

llvm-svn: 165239

11 years agoAdd a comment to the commit r165187.
Jakub Staszak [Thu, 4 Oct 2012 19:08:30 +0000 (19:08 +0000)]
Add a comment to the commit r165187.

llvm-svn: 165238

11 years ago- add tokens to PPCInstrInfo.td and PPCInstr64Bit.td to resolve
Will Schmidt [Thu, 4 Oct 2012 18:14:28 +0000 (18:14 +0000)]
- add tokens to PPCInstrInfo.td and PPCInstr64Bit.td to resolve
"Instruction 'foo' has no tokens" errors during llvm-tblgen
-gen-asm-matcher attempts.  At this time, the added
tokens are "#comment" style rather than the actual mnemonic.  This will
be revisited once the rest of the base asmparser bits get straightened
out for ppc64-elf-linux.

llvm-svn: 165237

11 years agoGet MCSchedModel directly from the subtarget.
Jakob Stoklund Olesen [Thu, 4 Oct 2012 17:30:43 +0000 (17:30 +0000)]
Get MCSchedModel directly from the subtarget.

Not all targets have itineraries, but the subtarget always has an
MCSchedModel.

llvm-svn: 165236

11 years agoSwitch MachineTraceMetrics to the new TargetSchedModel interface.
Jakob Stoklund Olesen [Thu, 4 Oct 2012 17:30:40 +0000 (17:30 +0000)]
Switch MachineTraceMetrics to the new TargetSchedModel interface.

llvm-svn: 165235

11 years agoFix scope location when parsing GNU attributes.
Michael Han [Thu, 4 Oct 2012 16:42:52 +0000 (16:42 +0000)]
Fix scope location when parsing GNU attributes.

For GNU attributes, instead of reusing attribute source
location for the scope location, use SourceLocation() since
GNU attributes don not have scope tokens.

llvm-svn: 165234

11 years agotest commit / whitespace
Will Schmidt [Thu, 4 Oct 2012 16:20:24 +0000 (16:20 +0000)]
test commit / whitespace

llvm-svn: 165233

11 years agoSimplifyCFG: Enhance the "remove CFG edge that leads to null pointer dereference...
Benjamin Kramer [Thu, 4 Oct 2012 16:11:49 +0000 (16:11 +0000)]
SimplifyCFG: Enhance the "remove CFG edge that leads to null pointer dereference" optimization to also handle instructions with multiple uses.

We conservatively only check the first use to avoid walking long use chains.
This catches the common case of having both a load and a store to a pointer
supplied by a PHI node.

llvm-svn: 165232

11 years agoRevert 165129
Preston Gurd [Thu, 4 Oct 2012 15:26:04 +0000 (15:26 +0000)]
Revert 165129

llvm-svn: 165231

11 years agotsan for Go: support mallocs before __tsan_init() (required to support cgo code)
Dmitry Vyukov [Thu, 4 Oct 2012 13:54:49 +0000 (13:54 +0000)]
tsan for Go: support mallocs before __tsan_init() (required to support cgo code)

llvm-svn: 165229

11 years agoIn my recent change to avoid use of underaligned memory I didn't notice that
Duncan Sands [Thu, 4 Oct 2012 13:53:21 +0000 (13:53 +0000)]
In my recent change to avoid use of underaligned memory I didn't notice that
cpyDest can be mutated in some cases, which would then cause a crash later if
indeed the memory was underaligned.  This brought down several buildbots, so
I guess the underaligned case is much more common than I thought!

llvm-svn: 165228

11 years agoAvoid calling __asan_init from memcpy(), memmove(), memset() during libSystem initial...
Alexander Potapenko [Thu, 4 Oct 2012 13:41:28 +0000 (13:41 +0000)]
Avoid calling __asan_init from memcpy(), memmove(), memset() during libSystem initialization.
Fixes http://code.google.com/p/address-sanitizer/issues/detail?id=117.

llvm-svn: 165227

11 years agoThe alignment of an sret parameter is known: it must be at least the
Duncan Sands [Thu, 4 Oct 2012 13:36:31 +0000 (13:36 +0000)]
The alignment of an sret parameter is known: it must be at least the
alignment of the return type.  Teach the optimizers this.

llvm-svn: 165226

11 years agoHoist some grossly duplicated code from the COFF/ELF/MachO streamers into MCObjectStr...
Benjamin Kramer [Thu, 4 Oct 2012 13:12:43 +0000 (13:12 +0000)]
Hoist some grossly duplicated code from the COFF/ELF/MachO streamers into MCObjectStreamer.

llvm-svn: 165225

11 years agoFix wrong name in comment.
Duncan Sands [Thu, 4 Oct 2012 13:07:26 +0000 (13:07 +0000)]
Fix wrong name in comment.

llvm-svn: 165224

11 years agoFix PR13969, a mini-phase-ordering issue with the new SROA pass.
Chandler Carruth [Thu, 4 Oct 2012 12:33:50 +0000 (12:33 +0000)]
Fix PR13969, a mini-phase-ordering issue with the new SROA pass.

Currently, we re-visit allocas when something changes about the way they
might be *split* to allow better scalarization to take place. However,
we weren't handling the case when the *promotion* is what would change
the behavior of SROA. When an address derived from an alloca is stored
into another alloca, we consider the first to have escaped. If the
second is ever promoted to an SSA value, we will suddenly be able to run
the SROA pass on the first alloca.

This patch adds explicit support for this form if iteration. When we
detect a store of a pointer derived from an alloca, we flag the
underlying alloca for reprocessing after promotion. The logic works hard
to only do this when there is definitely going to be promotion and it
might remove impediments to the analysis of the alloca.

Thanks to Nick for the great test case and Benjamin for some sanity
check review.

llvm-svn: 165223

11 years ago[ASan] don't build 64-bit runtime on 32-bit Linux platforms
Alexey Samsonov [Thu, 4 Oct 2012 12:22:33 +0000 (12:22 +0000)]
[ASan] don't build 64-bit runtime on 32-bit Linux platforms

llvm-svn: 165222

11 years agoThe memcpy optimizer was happily doing call slot forwarding when the new memory
Duncan Sands [Thu, 4 Oct 2012 10:54:40 +0000 (10:54 +0000)]
The memcpy optimizer was happily doing call slot forwarding when the new memory
was less aligned than the old.  In the testcase this results in an overaligned
memset: the memset alignment was correct for the original memory but is too much
for the new memory.  Fix this by either increasing the alignment of the new
memory or bailing out if that isn't possible.  Should fix the gcc-4.7 self-host
buildbot failure.

llvm-svn: 165220

11 years agoTeach the integer-promotion rewrite strategy to be endianness aware.
Chandler Carruth [Thu, 4 Oct 2012 10:39:28 +0000 (10:39 +0000)]
Teach the integer-promotion rewrite strategy to be endianness aware.
Sorry for this being broken so long. =/

As part of this, switch all of the existing tests to be Little Endian,
which is the behavior I was asserting in them anyways! Add in a new
big-endian test that checks the interesting behavior there.

Another part of this is to tighten the rules abotu when we perform the
full-integer promotion. This logic now rejects cases where there fully
promoted integer is a non-multiple-of-8 bitwidth or cases where the
loads or stores touch bits which are in the allocated space of the
alloca but are not loaded or stored when accessing the integer. Sadly,
these aren't really observable today as the rest of the pass will
already ensure the invariants hold. However, the latter situation is
likely to become a potential concern in the future.

Thanks to Benjamin and Duncan for early review of this patch. I'm still
looking into whether there are further endianness issues, please let me
know if anyone sees BE failures persisting past this.

llvm-svn: 165219

11 years agotsan: add atomic_fetch_sub() and atomic_signal_fence() functions
Dmitry Vyukov [Thu, 4 Oct 2012 10:08:23 +0000 (10:08 +0000)]
tsan: add atomic_fetch_sub() and atomic_signal_fence() functions

llvm-svn: 165218

11 years agoPrefer StringRef::startswith to the strncmp/strlen contraption.
Benjamin Kramer [Thu, 4 Oct 2012 10:06:38 +0000 (10:06 +0000)]
Prefer StringRef::startswith to the strncmp/strlen contraption.

This may be slightly more efficient and is definitely more readable.

llvm-svn: 165217

11 years agoFix typo in comments.
Logan Chien [Thu, 4 Oct 2012 08:08:56 +0000 (08:08 +0000)]
Fix typo in comments.

llvm-svn: 165216

11 years agoFix r165005: The lexical DeclContext is not the right place to make a decision about...
Axel Naumann [Thu, 4 Oct 2012 07:34:43 +0000 (07:34 +0000)]
Fix r165005: The lexical DeclContext is not the right place to make a decision about whether we need to call tryAddTopLevelDecl or not. That call should be made when the DeclContext's redeclaration context is the translation unit.

llvm-svn: 165215

11 years ago[asan] properly report mmap failure
Kostya Serebryany [Thu, 4 Oct 2012 07:21:09 +0000 (07:21 +0000)]
[asan] properly report mmap failure

llvm-svn: 165214

11 years agoUse method to query if there are attributes.
Bill Wendling [Thu, 4 Oct 2012 07:19:46 +0000 (07:19 +0000)]
Use method to query if there are attributes.

llvm-svn: 165213

11 years agoAdd method to query for NoCapture attribute.
Bill Wendling [Thu, 4 Oct 2012 07:18:12 +0000 (07:18 +0000)]
Add method to query for NoCapture attribute.

llvm-svn: 165212

11 years agoUse method to query for NoAlias attribute.
Bill Wendling [Thu, 4 Oct 2012 07:17:46 +0000 (07:17 +0000)]
Use method to query for NoAlias attribute.

llvm-svn: 165211

11 years agoUse attribute query methods.
Bill Wendling [Thu, 4 Oct 2012 07:08:30 +0000 (07:08 +0000)]
Use attribute query methods.

llvm-svn: 165210

11 years agoUse method to query for attributes.
Bill Wendling [Thu, 4 Oct 2012 06:58:52 +0000 (06:58 +0000)]
Use method to query for attributes.

llvm-svn: 165209

11 years agoAdd method to query for 'NoAlias' attribute on call/invoke instructions.
Bill Wendling [Thu, 4 Oct 2012 06:52:09 +0000 (06:52 +0000)]
Add method to query for 'NoAlias' attribute on call/invoke instructions.

llvm-svn: 165208

11 years agoUse method to query for attributes.
Bill Wendling [Thu, 4 Oct 2012 06:49:41 +0000 (06:49 +0000)]
Use method to query for attributes.

llvm-svn: 165207

11 years agoQuery for attributes via the correct method call.
Bill Wendling [Thu, 4 Oct 2012 06:48:57 +0000 (06:48 +0000)]
Query for attributes via the correct method call.

llvm-svn: 165206

11 years agoUse new accessor methods to query for attributes.
Bill Wendling [Thu, 4 Oct 2012 06:43:21 +0000 (06:43 +0000)]
Use new accessor methods to query for attributes.

llvm-svn: 165205

11 years ago[tsan] add 3 internal flags for fine-grain control of what is instrumented and what...
Kostya Serebryany [Thu, 4 Oct 2012 05:28:50 +0000 (05:28 +0000)]
[tsan] add 3 internal flags for fine-grain control of what is instrumented and what is not.

llvm-svn: 165204

11 years agoRemove template from function that is only used with one type after r165092.
Craig Topper [Thu, 4 Oct 2012 05:18:31 +0000 (05:18 +0000)]
Remove template from function that is only used with one type after r165092.

llvm-svn: 165203

11 years agoAdd missing comment for mangling.
Nick Lewycky [Thu, 4 Oct 2012 04:58:17 +0000 (04:58 +0000)]
Add missing comment for mangling.

llvm-svn: 165202

11 years agoFix reg mask slot test, and preserve LiveIntervals and VirtRegMap in the PBQP
Lang Hames [Thu, 4 Oct 2012 04:50:53 +0000 (04:50 +0000)]
Fix reg mask slot test, and preserve LiveIntervals and VirtRegMap in the PBQP
allocator. Fixes PR13945.

llvm-svn: 165201

11 years agodocs: Fix typo on front page
Sean Silva [Thu, 4 Oct 2012 04:41:27 +0000 (04:41 +0000)]
docs: Fix typo on front page

llvm-svn: 165200

11 years agoImplement methods that enable expansion of load immediate
Jack Carter [Thu, 4 Oct 2012 04:03:53 +0000 (04:03 +0000)]
Implement methods that enable expansion of load immediate
macro instruction (li) in the assembler.

We have identified three possible expansions depending on
the size of immediate operand:
  1) for 0 ≤ j ≤ 65535.
     li d,j =>
     ori d,$zero,j

  2) for −32768 ≤ j < 0.
     li d,j =>
     addiu d,$zero,j

  3) for any other value of j that is representable as a 32-bit integer.
     li d,j =>
     lui d,hi16(j)
     ori d,d,lo16(j)

All of the above have been implemented in ths patch.

Contributer: Vladimir Medic
llvm-svn: 165199

11 years agodocs: Sphinxify GoldPlugin document.
Sean Silva [Thu, 4 Oct 2012 03:56:23 +0000 (03:56 +0000)]
docs: Sphinxify GoldPlugin document.

llvm-svn: 165198

11 years agoFail early with a clear assert if an operation with multiple uses somehow ends
Lang Hames [Thu, 4 Oct 2012 03:23:25 +0000 (03:23 +0000)]
Fail early with a clear assert if an operation with multiple uses somehow ends
up being contracted during codegen.

llvm-svn: 165197

11 years agoTest file-scoped FP_CONTRACT pragma.
Lang Hames [Thu, 4 Oct 2012 03:07:37 +0000 (03:07 +0000)]
Test file-scoped FP_CONTRACT pragma.

llvm-svn: 165196

11 years agoPermanently end the whole "pragma got handled by the parser too early"
Eli Friedman [Thu, 4 Oct 2012 02:36:51 +0000 (02:36 +0000)]
Permanently end the whole "pragma got handled by the parser too early"
mess by handling all pragmas which the parser touches uniformly.
<rdar://problem/12248901>, etc.

llvm-svn: 165195

11 years agoThis patch is a partial implementation of mips .set assembler directive. Directive...
Jack Carter [Thu, 4 Oct 2012 02:29:46 +0000 (02:29 +0000)]
This patch is a partial implementation of mips .set assembler directive. Directive is defined as follows:
.set option
The patch implements following options

    at - lets the assembler use the $at register for macros,
         but generates warnings if the source program uses $at

    noat - let source programs use $at without issuingwarnings.

    noreorder - prevents the assembler from reordering machine
                language instructions.
    nomacro - causes the assembler to print a warning whenever
              an assembler operation generates more than one
              machine language instruction.
    macro - lets the assembler generate multiple machine instructions
            from a single assembler instruction
    reorder - lets the assembler reorder machine language
               instructions to improve performance

The above variants are parsed and their boolean values set or unset.
The code to actually use them will come later.

Following options are not implemented yet:

nomips16
nomicromips
move
nomove

Contributer: Vladimir Medic
llvm-svn: 165194

11 years agoPatch submitted by Dan Malea -- I introduced a dependency between
Jason Molenda [Thu, 4 Oct 2012 02:16:06 +0000 (02:16 +0000)]
Patch submitted by Dan Malea -- I introduced a dependency between
ProcessGDBRemote and DynamicLoaderDarwinKernel and a patch was needed
to get this building on Linux.  Thanks!

llvm-svn: 165193