Don't create VBROADCAST nodes if any nodes use the chain result from the load. Fixes PR11900.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149478 91177308-0d34-0410-b5e6-96231b3b80d8

BBVectorize.cpp: Try to fix MSVC build. map::iterator and multimap::iterator are incompatible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149475 91177308-0d34-0410-b5e6-96231b3b80d8

A few of the changes suggested in code review (by Nick Lewycky)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149472 91177308-0d34-0410-b5e6-96231b3b80d8

Revert Chris' commits up to r149348 that started causing VMCoreTests unit test to fail.

These are:

r149348
r149351
r149352
r149354
r149356
r149357
r149361
r149362
r149364
r149365

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149470 91177308-0d34-0410-b5e6-96231b3b80d8

Add a basic-block autovectorization pass.

This is the initial checkin of the basic-block autovectorization pass along with some supporting vectorization infrastructure.
Special thanks to everyone who helped review this code over the last several months (especially Tobias Grosser).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149468 91177308-0d34-0410-b5e6-96231b3b80d8

Disable InstCombine unsafe folding bitcasts of calls w/ varargs.

Changing arguments from being passed as fixed to varargs is unsafe, as
the ABI may require they be handled differently (stack vs. register, for
example).

Remove two tests which rely on the bitcast being folded into the direct
call, which is exactly the transformation that's unsafe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149457 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up. One more return type mismatch fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149452 91177308-0d34-0410-b5e6-96231b3b80d8

Revert test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149446 91177308-0d34-0410-b5e6-96231b3b80d8

bz11794 : EarlyCSE stack overflow on long functions.
Make the EarlyCSE optimizer not use recursion to do a depth first iteration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149445 91177308-0d34-0410-b5e6-96231b3b80d8

test commit, adding a blank space

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149444 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a crash in llvm-mc for Mach-O when a symbol difference expression uses a
symbol from an assignment.  In this case the symbol did not have a fragment so
MCObjectWriter::IsSymbolRefDifferenceFullyResolved() should not have been
calling IsSymbolRefDifferenceFullyResolvedImpl() with a NULL fragment and should
just have returned false in that case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149442 91177308-0d34-0410-b5e6-96231b3b80d8

Add Triple::getMacOSXVersion to replace crufty code in the clang driver.

This new function provides a way to get the Mac OS X version number from
either generic "darwin" triples of macosx triples.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149438 91177308-0d34-0410-b5e6-96231b3b80d8

Don't assign a value to NUM_TARGET_NAMED_SUBREGS.

It was wrong and completely unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149433 91177308-0d34-0410-b5e6-96231b3b80d8

Move the composite map into CodeGenSubRegIndex.

Each SubRegIndex keeps track of how it composes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149423 91177308-0d34-0410-b5e6-96231b3b80d8

Add a TableGen CodeGenSubRegIndex class.

This class is used to represent SubRegIndex instances instead of the raw
Record pointers that were used before.

No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149418 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor loop for better readability.

Excellent suggestion from Ben Kramer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149417 91177308-0d34-0410-b5e6-96231b3b80d8

Add explanatory comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149416 91177308-0d34-0410-b5e6-96231b3b80d8

RefCountedBaseVPTR needs the IntrusiveRefCntPtrInfo as friend,
now that this handles the release / retain calls.

Adds a regression test for that bug (which is a compile-time
regression) and for the last two changes to the IntrusiveRefCntPtr,
especially tests for the memory leak due to copy construction of the
ref-counted object and ensuring that the traits are used for release /
retain calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149411 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up. Trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149408 91177308-0d34-0410-b5e6-96231b3b80d8

Obvious unnecessary loop removal. Follow through from previous checkin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149398 91177308-0d34-0410-b5e6-96231b3b80d8

Add assembler dialect attribute in asm parser which lets target specific asm parser change dialect on the fly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149396 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149391 91177308-0d34-0410-b5e6-96231b3b80d8

Fix BrainF compilation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149375 91177308-0d34-0410-b5e6-96231b3b80d8

Increase the initial vector size to be equivalent to the size of the Deps
vector. This potentially saves a resizing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149369 91177308-0d34-0410-b5e6-96231b3b80d8

Cache the size of the vector instead of calling .size() all over the place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149368 91177308-0d34-0410-b5e6-96231b3b80d8

Remove pcmpgt/pcmpeq intrinsics as clang is not using them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149367 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the "string" form of ConstantArray::get, using
ConstantDataArray::getString instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149365 91177308-0d34-0410-b5e6-96231b3b80d8

remove commented-out code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149364 91177308-0d34-0410-b5e6-96231b3b80d8

with recent changes, ConstantArray is never a "string".  Remove the associated
methods and constant fold the clients to false.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149362 91177308-0d34-0410-b5e6-96231b3b80d8

update this to ConstantDataArray.  There are no tests and this isn't using the preferred functionality for ripping apart strings, so I have no way to test this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149361 91177308-0d34-0410-b5e6-96231b3b80d8

RAFast: Generalize the logic for return operands.

This removes implicit assumption about the form of MI coming into regalloc. In particular, it should be independent of ProcessImplicitDefs which will eventually become a standard part of coming out of SSA--unless we simply can eliminate IMPLICIT_DEF completely. Current unit tests expose this once I remove incidental pass ordering restrictions.

This is not a final fix. Just a temporary workaround until I figure out the right way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149360 91177308-0d34-0410-b5e6-96231b3b80d8

fix a small oversight that broke the fhourstones app.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149357 91177308-0d34-0410-b5e6-96231b3b80d8

remove the last vestiges of llvm::GetConstantStringInfo, in CodeGen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149356 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the last uses of GetConstantStringInfo from this file, I didn't realize I was that close...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149354 91177308-0d34-0410-b5e6-96231b3b80d8

Add various coarse bit-width architecture predicates to llvm::Triple.
These are very useful for frontends and other utilities reasoning about
or selecting between triples.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149353 91177308-0d34-0410-b5e6-96231b3b80d8

start moving SimplifyLibcalls over to getConstantStringInfo, which is
dramatically more efficient than GetConstantStringInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149352 91177308-0d34-0410-b5e6-96231b3b80d8

Change ConstantArray::get to form a ConstantDataArray when possible,
kicking in the big win of ConstantDataArray.  As part of this, change
the implementation of GetConstantStringInfo in ValueTracking to work
with ConstantDataArray (and not ConstantArray) making it dramatically,
amazingly, more efficient in the process and renaming it to
getConstantStringInfo.

This keeps around a GetConstantStringInfo entrypoint that (grossly)
forwards to getConstantStringInfo and constructs the std::string
required, but existing clients should move over to
getConstantStringInfo instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149351 91177308-0d34-0410-b5e6-96231b3b80d8

rework this logic to not depend on the last argument to GetConstantStringInfo,
which is going away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149348 91177308-0d34-0410-b5e6-96231b3b80d8

don't emit a 1-byte object as a .fill.  This is silly and causes
CodeGen/X86/global-sections.ll to fail with CDArray

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149343 91177308-0d34-0410-b5e6-96231b3b80d8

use the right accessor for ConstantDataArray.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149342 91177308-0d34-0410-b5e6-96231b3b80d8

fix asmwriting of ConstantDataArray to use the right element count,
simplify ConstantArray handling, since they can never be empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149341 91177308-0d34-0410-b5e6-96231b3b80d8

enhance logic to support ConstantDataArray.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149340 91177308-0d34-0410-b5e6-96231b3b80d8

Remove all references to the old EH.

There was always the current EH. -- Ministry of Truth

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149335 91177308-0d34-0410-b5e6-96231b3b80d8

Update test to new EH model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149333 91177308-0d34-0410-b5e6-96231b3b80d8

Update test to new EH model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149332 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the now-dead llvm.eh.exception and llvm.eh.selector intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149331 91177308-0d34-0410-b5e6-96231b3b80d8

s/getInnerUnwindDest/getInnerResumeDest/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149328 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the eh.exception and eh.selector intrinsics. Also remove a hack to copy
over the catch information. The catch information is now tacked to the invoke
instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149326 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ivar which is identical to another ivar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149323 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused ivars and s/getOuterUnwindDest/getOuterResumeDest/g.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149322 91177308-0d34-0410-b5e6-96231b3b80d8

Remove more dead functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149318 91177308-0d34-0410-b5e6-96231b3b80d8

s/getInnerUnwindDestNewEH/getInnerUnwindDest/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149317 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some unused, old-EH methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149316 91177308-0d34-0410-b5e6-96231b3b80d8

Use the correct ShiftAmtTy for creating shifts after legalization.  PR11881.  Not committing a testcase because I think it will be too fragile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149315 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of references to dead intrinsics.

The eh.selector and eh.resume intrinsics aren't used anymore. Get rid of some
calls to them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149314 91177308-0d34-0410-b5e6-96231b3b80d8

Formatting cleanups. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149312 91177308-0d34-0410-b5e6-96231b3b80d8

Relax constructor for IntrusiveRefCntPtr to not be explicit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149309 91177308-0d34-0410-b5e6-96231b3b80d8

Use traits for IntrusiveRefCntPtr to determine how to increment/decrement a reference count.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149308 91177308-0d34-0410-b5e6-96231b3b80d8

Remove no-longer-useful dyn_casts and pals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149307 91177308-0d34-0410-b5e6-96231b3b80d8

Add a constified getLandingPad() method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149303 91177308-0d34-0410-b5e6-96231b3b80d8

[asan] fix the ObjC support (asan Issue #33)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149300 91177308-0d34-0410-b5e6-96231b3b80d8

Chris's constant data sequence refactoring actually enabled printing
vectors of all one bits to be printed more cleverly in the AsmPrinter.
Unfortunately, the byte value for all one bits is the same with
-fsigned-char as the error return of '-1'. Force this to be the unsigned
byte value when returning it to avoid this problem, and update the test
case for the shiny new behavior.

Yay for building LLVM and Clang with -funsigned-char.

Chris, please review, and let me know if there is any reason to not
desire this change. It seems good on the surface, and certainly intended
based on the code written.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149299 91177308-0d34-0410-b5e6-96231b3b80d8

PR11834: Use macros which are defined on Windows. Patch by Marina Yatsina.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149294 91177308-0d34-0410-b5e6-96231b3b80d8

basic-aa does support AliasAnalysis chaining now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149293 91177308-0d34-0410-b5e6-96231b3b80d8

Intel syntax. Adjust special code, used to recognize cmp<comparison code>{ss,sd,ps,pd}, for intel syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149291 91177308-0d34-0410-b5e6-96231b3b80d8

Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149289 91177308-0d34-0410-b5e6-96231b3b80d8

Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149275 91177308-0d34-0410-b5e6-96231b3b80d8

Implement String Cast from Integer

Allow casts from integer to string.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149273 91177308-0d34-0410-b5e6-96231b3b80d8

Intel syntax. Support .intel_syntax directive.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149270 91177308-0d34-0410-b5e6-96231b3b80d8

Fix refacto.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149269 91177308-0d34-0410-b5e6-96231b3b80d8

Here's a new one: GCC was complaining about an only-used-in-asserts
*function*. Wrap the function in #ifndef NDEBUG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149259 91177308-0d34-0410-b5e6-96231b3b80d8

Various improvements suggested by Duncan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149255 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate narrowing conversion in initializer list, to make C++11 happy

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149254 91177308-0d34-0410-b5e6-96231b3b80d8

X86: Simplify shuffle mask generation code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149248 91177308-0d34-0410-b5e6-96231b3b80d8

Fix compilation of ASan tests on OS X Lion (see code.google.com/p/address-sanitizer/issues/detail?id=32)
The redzones emitted by AddressSanitizer for CFString instances confuse the linker and are of little use, so we shouldn't add them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149243 91177308-0d34-0410-b5e6-96231b3b80d8

Add GCCBuiltin declarations for cmpsd/cmpss/cmppd/cmpps to allow custom code to be removed from clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149234 91177308-0d34-0410-b5e6-96231b3b80d8

Fix pattern for memory form of PSHUFD for use with FP vectors to remove bitcast to an integer vector that normal code wouldn't have. Also remove bitcasts from code that turns splat vector loads into a shuffle as it was making the broken pattern necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149232 91177308-0d34-0410-b5e6-96231b3b80d8

fix a major oversight that is breaking some llvm-test tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149230 91177308-0d34-0410-b5e6-96231b3b80d8

DenseMap::find_as() and unit tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149229 91177308-0d34-0410-b5e6-96231b3b80d8

First step of flipping on ConstantDataSequential: enable ConstantDataVector
to be formed whenever ConstantVector::get is used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149226 91177308-0d34-0410-b5e6-96231b3b80d8

when verbose asm is on, print integers in ConstantDataSequentials just
like normal integers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149223 91177308-0d34-0410-b5e6-96231b3b80d8

don't lose tail padding on ConstantDataAggregate vec3's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149222 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ConstantFoldShuffleVectorInstruction to properly handle the case
when the result type has a different # elements than the input vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149221 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Promote the testing targets out of folders on IDE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149220 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Add the folder "Tablegenning" on llvm-tblgen targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149219 91177308-0d34-0410-b5e6-96231b3b80d8

initialize the Next field to null

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149217 91177308-0d34-0410-b5e6-96231b3b80d8

Move some XOP patterns into instruction definition. Replae VPCMOV intrinsic patterns with custom lowering to a target specific nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149216 91177308-0d34-0410-b5e6-96231b3b80d8

Add bitcode reader and writer support for ConstantDataAggregate, which
should be feature complete now.  Lets see if it works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149215 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup the organization of some methods in llvm::Triple and provide
a better doxyment group for convenience predicates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149211 91177308-0d34-0410-b5e6-96231b3b80d8

Move Clang's file-level locking facility over to LLVM's support
library, since it doesn't really have anything to do with Clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149203 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanups for EABI standard functions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149195 91177308-0d34-0410-b5e6-96231b3b80d8

Use base AAPCS for varargs functions even for AAPCS-VFP CC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149194 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some scavenger performance issues.

- Don't call malloc+free in the very hot forward().
- Don't call isTiedToDefOperand().
- Don't create BitVector temporaries.
- Merge DeadRegs into KillRegs.
- Eliminate the early clobber checks, they were irrelevant to scavenging.
- Remove unnecessary code from -Asserts builds.

This speeds up ARM PEI by 3.4x and overall llc -O0 codegen time by 11%.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149189 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid creating BitVector temporaries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149188 91177308-0d34-0410-b5e6-96231b3b80d8

Add a BitVector::reset(BitVector&) method.

The alternative LHS &= ~RHS is way too slow because it creates a
temporary that calls malloc/free.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149187 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149185 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note about a potential optimization for clz/ctz patterns for ARM
(and other targets).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149182 91177308-0d34-0410-b5e6-96231b3b80d8

Ensure .AliasedSymbol() is called on all uses of getSymbol(). Affects ARM and MIPS ELF backends.

Fixes PR11877

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149180 91177308-0d34-0410-b5e6-96231b3b80d8

Silence GCC's -Wreturn-type warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149179 91177308-0d34-0410-b5e6-96231b3b80d8

Small improvement to the recursion detection logic from the previous commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149175 91177308-0d34-0410-b5e6-96231b3b80d8