oota-llvm.git
9 years ago[ValueTracking] Remove untested / unreachable code, NFC
Sanjoy Das [Wed, 11 Nov 2015 00:16:41 +0000 (00:16 +0000)]
[ValueTracking] Remove untested / unreachable code, NFC

Right now isTruePredicate is only ever called with Pred == ICMP_SLE or
ICMP_ULE, and the ICMP_SLT and ICMP_ULT cases are dead.  This change
removes the untested dead code so that the function is not misleading.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252676 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAMDGPU: Remove dead code
Matt Arsenault [Wed, 11 Nov 2015 00:01:36 +0000 (00:01 +0000)]
AMDGPU: Remove dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252675 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAMDGPU: Set isAllocatable = 0 on VS_32/VS_64
Matt Arsenault [Wed, 11 Nov 2015 00:01:32 +0000 (00:01 +0000)]
AMDGPU: Set isAllocatable = 0 on VS_32/VS_64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252674 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ValueTracking] Teach isImpliedCondition a new bitwise trick
Sanjoy Das [Tue, 10 Nov 2015 23:56:20 +0000 (23:56 +0000)]
[ValueTracking] Teach isImpliedCondition a new bitwise trick

Summary:
This change teaches isImpliedCondition to prove things like

  (A | 15) < L  ==>  (A | 14) < L

if the low 4 bits of A are known to be zero.

Depends on D14391

Reviewers: majnemer, reames, hfinkel

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14392

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252673 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ValueTracking] Use m_APInt instead of m_ConstantInt, NFC
Sanjoy Das [Tue, 10 Nov 2015 23:56:15 +0000 (23:56 +0000)]
[ValueTracking] Use m_APInt instead of m_ConstantInt, NFC

This change would add functionality if isImpliedCondition worked on
vector types; but since it bail out on vector predicates this change is
an NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252672 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTableGen: Emit LaneMask for register classes without subregisters as ~0u
Matthias Braun [Tue, 10 Nov 2015 23:23:05 +0000 (23:23 +0000)]
TableGen: Emit LaneMask for register classes without subregisters as ~0u

This makes it slightly easier to handle classes with and without
subregister uniformly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252671 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WinEH] Insert the MBB for EH_RESTORE after the catchret
Reid Kleckner [Tue, 10 Nov 2015 23:22:20 +0000 (23:22 +0000)]
[WinEH] Insert the MBB for EH_RESTORE after the catchret

Inserting it before the target block could be bad, we might already have
a fallthrough edge to it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252670 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[cmake] move SONAME handling to llvm_add_library
Andrew Wilkins [Tue, 10 Nov 2015 23:19:21 +0000 (23:19 +0000)]
[cmake] move SONAME handling to llvm_add_library

Summary:
Move handling of the SONAME option from add_llvm_library
to llvm_add_library, so that it can be used in sub-projects.
In particular, this makes it possible to have consistently
named shared libraries for LLVM, Clang and LLDB.

Also, base the SONAME and symlinks on the output name
by extracting the OUTPUT_NAME property, rather than assuming
it is the same as the target name.

Reviewers: beanz

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14539

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252669 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agodon't repeat function/class/variable names in comments; NFC
Sanjay Patel [Tue, 10 Nov 2015 22:59:11 +0000 (22:59 +0000)]
don't repeat function/class/variable names in comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252666 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[doc] Compile CUDA with LLVM
Jingyue Wu [Tue, 10 Nov 2015 22:35:47 +0000 (22:35 +0000)]
[doc] Compile CUDA with LLVM

Summary:
This patch adds documentation on compiling CUDA with LLVM as requested by many
engineers and researchers. It includes not only user guides but also some
internals (mostly optimizations) so that early adopters can start hacking and
contributing.

Quite a few researchers who contacted us haven't used LLVM before, which is
unsurprising as it hasn't been long since LLVM picked up CUDA. So I added a
short summary to help these folks get started with LLVM.

I expect this document to evolve substantially down the road. The user guides
will be much simplified after the Clang integration is done. However, the
internals should continue growing to include for example performance debugging
and key areas to improve.

Reviewers: chandlerc, meheff, broune, tra

Subscribers: silvas, jingyue, llvm-commits, eliben

Differential Revision: http://reviews.llvm.org/D14370

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252660 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[libFuzzer] add UninstrumentedTest.cpp (missing from a previous commit)
Kostya Serebryany [Tue, 10 Nov 2015 22:02:56 +0000 (22:02 +0000)]
[libFuzzer] add UninstrumentedTest.cpp (missing from a previous commit)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252658 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WebAssembly] Remove special cases for things that are no longer special. NFC.
Dan Gohman [Tue, 10 Nov 2015 21:48:21 +0000 (21:48 +0000)]
[WebAssembly] Remove special cases for things that are no longer special. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252656 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd PPCMIPeephole.cpp to CMakeLists.txt
Bill Schmidt [Tue, 10 Nov 2015 21:43:45 +0000 (21:43 +0000)]
Add PPCMIPeephole.cpp to CMakeLists.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252654 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WebAssembly] Support for floating point min and max.
Dan Gohman [Tue, 10 Nov 2015 21:40:21 +0000 (21:40 +0000)]
[WebAssembly] Support for floating point min and max.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252653 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[CMake] Autoconf builds libLTO with -fPIC, CMake should be able to as well.
Chris Bieneman [Tue, 10 Nov 2015 21:38:58 +0000 (21:38 +0000)]
[CMake] Autoconf builds libLTO with -fPIC, CMake should be able to as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252652 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[PowerPC] Add an MI SSA peephole pass.
Bill Schmidt [Tue, 10 Nov 2015 21:38:26 +0000 (21:38 +0000)]
[PowerPC] Add an MI SSA peephole pass.

This patch adds a pass for doing PowerPC peephole optimizations at the
MI level while the code is still in SSA form.  This allows for easy
modifications to the instructions while depending on a subsequent pass
of DCE.  Both passes are very fast due to the characteristics of SSA.

At this time, the only peepholes added are for cleaning up various
redundancies involving the XXPERMDI instruction.  However, I would
expect this will be a useful place to add more peepholes for
inefficiencies generated during instruction selection.  The pass is
placed after VSX swap optimization, as it is best to let that pass
remove unnecessary swaps before performing any remaining clean-ups.

The utility of these clean-ups are demonstrated by changes to four
existing test cases, all of which now have tighter expected code
generation.  I've also added Eric Schweiz's bugpoint-reduced test from
PR25157, for which we now generate tight code.  One other test started
failing for me, and I've fixed it
(test/Transforms/PlaceSafepoints/finite-loops.ll) as well; this is not
related to my changes, and I'm not sure why it works before and not
after.  The problem is that the CHECK-NOT: of "statepoint" from test1
fails because of the "statepoint" in test2, and so forth.  Adding a
CHECK-LABEL in between keeps the different occurrences of that string
properly scoped.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252651 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agodsymutil: Prune module forward decl DIEs if a uniquable definition was
Adrian Prantl [Tue, 10 Nov 2015 21:31:05 +0000 (21:31 +0000)]
dsymutil: Prune module forward decl DIEs if a uniquable definition was
already emitted and fix a latent bug in DIECloner where the DW_CHILDREN_yes
flag is set based on the number of children in the input DIE rather than
the number of children that are actually being cloned.

rdar://problem/23439845

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252649 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoEnsure ModuleLinker materializes complete comdat groups
Teresa Johnson [Tue, 10 Nov 2015 21:09:06 +0000 (21:09 +0000)]
Ensure ModuleLinker materializes complete comdat groups

Summary:
The module linker lazy links some "discardable if unused" global
values (e.g. linkonce), materializing and linking them only
if they are referenced in the module. If a comdat group contains a
linkonce member that is not referenced, however, it would not be
materialized and linked, leading to an incomplete comdat group.

If there are other object files not part of the same LTO link that also
define and use that comdat group, the linker may select the incomplete
group leading to link time unsats.

To solve this, whenever a global value body is linked, make sure we
materialize any other members of the same comdat group that are not yet
materialized. This ensures they are in the lazy link list and get linked
as well.

Added new test and adjusted old test to remove parts that didn't
make sense with fix.

Reviewers: rafael

Subscribers: dexonsmith, davidxl, llvm-commits

Differential Revision: http://reviews.llvm.org/D14516

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252647 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[IR] Make {Call,Invoke}::cloneImpl aware of operand bundles
Sanjoy Das [Tue, 10 Nov 2015 20:13:21 +0000 (20:13 +0000)]
[IR] Make {Call,Invoke}::cloneImpl aware of operand bundles

This was an omission in the patch that landed initial support for
operand bundles.  So far we haven't hit this, but we will once the
inliner is able to inline calls to functions that contain calls with
operand bundles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252645 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[OperandBundles] Identify operand bundles with both their names and IDs
Sanjoy Das [Tue, 10 Nov 2015 20:13:15 +0000 (20:13 +0000)]
[OperandBundles] Identify operand bundles with both their names and IDs

No code uses this functionality yet.  This change just exposes
information / structure that was already present.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252644 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoless indent; NFCI
Sanjay Patel [Tue, 10 Nov 2015 20:09:02 +0000 (20:09 +0000)]
less indent; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252643 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ARM] add overrides for isCheapToSpeculateCttz() and isCheapToSpeculateCtlz()
Sanjay Patel [Tue, 10 Nov 2015 19:24:31 +0000 (19:24 +0000)]
[ARM] add overrides for isCheapToSpeculateCttz() and isCheapToSpeculateCtlz()

ARM V6T2 has instructions for efficient count-leading/trailing-zeros, so this should be
considered a cheap operation (and therefore fair game for speculation) for any ARM V6T2
implementation.

The net result of allowing this speculation for the regression tests in this patch is
that we get this code:

ctlz:
  clz  r0, r0
  bx  lr
cttz:
  rbit  r0, r0
  clz  r0, r0
  bx  lr

Instead of:

ctlz:
  cmp  r0, #0
  moveq  r0, #32
  clzne  r0, r0
  bx  lr
cttz:
  cmp   r0, #0
  moveq  r0, #32
  rbitne  r0, r0
  clzne  r0, r0
  bx  lr

This will help solve a general speculation/despeculation problem noted in PR24818:
https://llvm.org/bugs/show_bug.cgi?id=24818

Differential Revision: http://reviews.llvm.org/D14469

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252639 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agollvm-lto: trivial spelling changes to distinguish custom diagnostic handler and
Yunzhong Gao [Tue, 10 Nov 2015 18:52:48 +0000 (18:52 +0000)]
llvm-lto: trivial spelling changes to distinguish custom diagnostic handler and
default diagnostic handler.

Differential Revision: http://reviews.llvm.org/D14520

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252633 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLegalizeDAG: Implement promote for scalar_to_vector
Matt Arsenault [Tue, 10 Nov 2015 18:48:11 +0000 (18:48 +0000)]
LegalizeDAG: Implement promote for scalar_to_vector

This allows avoiding the default Expand behavior which
introduces stack usage. Bitcast the scalar and replace
the missing elements with undef.

This is covered by existing tests and used by a future
commit which makes 64-bit vectors legal types on AMDGPU.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252632 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLegalizeDAG: Implement promote for insert_vector_elt
Matt Arsenault [Tue, 10 Nov 2015 18:48:08 +0000 (18:48 +0000)]
LegalizeDAG: Implement promote for insert_vector_elt

This is covered by existing tests and used by a future
commit which makes 64-bit vectors legal types on AMDGPU.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252631 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLegalizeDAG: Implement promote for extract_vector_elt
Matt Arsenault [Tue, 10 Nov 2015 18:48:04 +0000 (18:48 +0000)]
LegalizeDAG: Implement promote for extract_vector_elt

This is for AMDGPU to implement v2i64 extract as extract of
half of a v4i32.

This is covered by existing tests and used by a future
commit which makes 64-bit vectors legal types on AMDGPU.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252630 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ValueTracking] Recognize that and(x, add (x, -1)) clears the low bit
Philip Reames [Tue, 10 Nov 2015 18:46:14 +0000 (18:46 +0000)]
[ValueTracking] Recognize that and(x, add (x, -1)) clears the low bit

This is a cleaned up version of a patch by John Regehr with permission. Originally found via the souper tool.

If we add an odd number to x, then bitwise-and the result with x, we know that the low bit of the result must be zero. Either it was zero in x originally, or the add cleared it in the temporary value. As a result, one of the two values anded together must have the bit cleared.

Differential Revision: http://reviews.llvm.org/D14315

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252629 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAddLLVM: squelch CMP0007 by not adding empty elements to list
Ramkumar Ramachandra [Tue, 10 Nov 2015 18:26:34 +0000 (18:26 +0000)]
AddLLVM: squelch CMP0007 by not adding empty elements to list

When configuring various llvm projects that use AddLLVM.cmake, this warning is
emitted many times, flooding the screen:

Policy CMP0007 is not set: list command no longer ignores empty elements.

The fix is removing an extra semicolon.

Differential Revision: http://reviews.llvm.org/D14339

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252628 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ThinLTO] Update comment per change in WeakAny handling (NFC)
Teresa Johnson [Tue, 10 Nov 2015 18:26:31 +0000 (18:26 +0000)]
[ThinLTO] Update comment per change in WeakAny handling (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252627 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ThinLTO] WeakAny fixes/cleanup
Teresa Johnson [Tue, 10 Nov 2015 18:20:11 +0000 (18:20 +0000)]
[ThinLTO] WeakAny fixes/cleanup

Ensure WeakAny variables are imported as ExternalWeak declarations. To
handle WeakAny more consistently and fix this issue:

1) Update helper doImportAsDefinition to properly flag WeakAny variables
   and aliases as not importing defintions.

   Update callers of doImportAsDefinition to remove now redundant checks for
   WeakAny aliases, or ignore aliases, as appropriate.

2) Add any !doImportAsDefinition GVs to DoNotLinkFromSource set during
   linking of the GV prototype, where we usually add GVs to the
   DoNotLinkFromSource set for other reasons.

   Remove now unnecessary adding of WeakAny aliases to
   DoNotLinkFromSource set from copyGlobalAliasProto.

   Remove now unnecessary guard against linking non-imported function
   bodies from ModuleLinker::run.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252626 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] add overrides for isCheapToSpeculateCttz() and isCheapToSpeculateCtlz()
Sanjay Patel [Tue, 10 Nov 2015 18:11:37 +0000 (18:11 +0000)]
[AArch64] add overrides for isCheapToSpeculateCttz() and isCheapToSpeculateCtlz()

AArch64 has instructions for efficient count-leading/trailing-zeros, so this should be
considered a cheap operation (and therefore fair game for speculation) for any AArch64
implementation.

The net result of allowing this speculation for the regression tests in this
patch is that we get this code:

ctlz:
  clz  w0, w0
  ret

cttz:
  rbit  w8, w0
  clz  w0, w8
  ret

Instead of:

ctlz:
  cbz  w0, .LBB0_2
  clz  w0, w0
  ret
.LBB0_2:
  orr  w0, wzr, #0x20
  ret

cttz:
  cbz  w0, .LBB1_2
  rbit  w8, w0
  clz  w0, w8
  ret
.LBB1_2:
  orr  w0, wzr, #0x20
  ret

See D14469 for the larger motivation.

Differential Revision: http://reviews.llvm.org/D14505

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252625 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Strip metadata when speculatively hoisting instructions"
Renato Golin [Tue, 10 Nov 2015 18:01:16 +0000 (18:01 +0000)]
Revert "Strip metadata when speculatively hoisting instructions"

This reverts commit r252604, as it broke all ARM and AArch64 buildbots, as
well as some x86, et al.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252623 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86] Do not try to custom-lower sitofp/fptosi in soft-float mode
Michael Kuperstein [Tue, 10 Nov 2015 17:37:49 +0000 (17:37 +0000)]
[X86] Do not try to custom-lower sitofp/fptosi in soft-float mode

Differential Revision: http://reviews.llvm.org/D14495

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252621 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix asan warning (NFC)
Xinliang David Li [Tue, 10 Nov 2015 17:11:33 +0000 (17:11 +0000)]
Fix asan warning (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252617 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoadd 'MustReduceDepth' as an objective/cost-metric for the MachineCombiner
Sanjay Patel [Tue, 10 Nov 2015 16:48:53 +0000 (16:48 +0000)]
add 'MustReduceDepth' as an objective/cost-metric for the MachineCombiner

This is one of the problems noted in PR25016:
https://llvm.org/bugs/show_bug.cgi?id=25016
and:
http://lists.llvm.org/pipermail/llvm-dev/2015-October/090998.html

The spilling problem is independent and not addressed by this patch.

The MachineCombiner was doing reassociations that don't improve or even worsen the critical path.
This is caused by inclusion of the "slack" factor when calculating the critical path of the original
code sequence. If we don't add that, then we have a more conservative cost comparison of the old code
sequence vs. a new sequence. The more liberal calculation must be preserved, however, for the AArch64
MULADD patterns because benchmark regressions were observed without that.

The two failing test cases now have identical asm that does what we want:
a + b + c + d ---> (a + b) + (c + d)

Differential Revision: http://reviews.llvm.org/D13417

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252616 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReapply "[ARM] Combine CMOV into BFI where possible"
James Molloy [Tue, 10 Nov 2015 14:22:05 +0000 (14:22 +0000)]
Reapply "[ARM] Combine CMOV into BFI where possible"

Added fixes for stage2 failures: CMOV is not commutable; commuting the operands results in the condition being flipped! d'oh!

Original commit message:

If we have a CMOV, OR and AND combination such as:
  if (x & CN)
      y |= CM;

And:
  * CN is a single bit;
    * All bits covered by CM are known zero in y;

Then we can convert this to a sequence of BFI instructions. This will always be a win if CM is a single bit, will always be no worse than the TST & OR sequence if CM is two bits, and for thumb will be no worse if CM is three bits (due to the extra IT instruction).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252606 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoStrip metadata when speculatively hoisting instructions
Igor Laevsky [Tue, 10 Nov 2015 14:10:31 +0000 (14:10 +0000)]
Strip metadata when speculatively hoisting instructions

This is fix for PR24059.

When we are hoisting instruction above some condition it may turn out
that metadata on this instruction was control dependant on the condition.
This metadata becomes invalid and we need to drop it.

This patch should cover most obvious places of speculative execution (which
I have found by greping isSafeToSpeculativelyExecute). I think there are more
cases but at least this change covers the severe ones.

Differential Revision: http://reviews.llvm.org/D14398

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252604 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUpdate test to use explicit triple
Oliver Stannard [Tue, 10 Nov 2015 14:09:08 +0000 (14:09 +0000)]
Update test to use explicit triple

This is needed for targets which do not support big-endian with the default
triple.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252603 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[PowerPC] Remove redundant code.
Tilmann Scheller [Tue, 10 Nov 2015 12:29:37 +0000 (12:29 +0000)]
[PowerPC] Remove redundant code.

The local variable Hi is never being read.

Issue identified by the Clang static analyzer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252600 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Fix halfword load merging for big-endian targets
Oliver Stannard [Tue, 10 Nov 2015 11:04:18 +0000 (11:04 +0000)]
[AArch64] Fix halfword load merging for big-endian targets

For big-endian targets, when we merge two halfword loads into a word load, the
order of the halfwords in the loaded value is reversed compared to
little-endian, so the load-store optimiser needs to swap the destination
registers.

This does not affect merging of two word loads, as we use ldp, which treats the
memory as two separate 32-bit words.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252597 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInliner: Do zero-cost inlines even if above a negative threshold (PR24851)
Hans Wennborg [Tue, 10 Nov 2015 09:47:48 +0000 (09:47 +0000)]
Inliner: Do zero-cost inlines even if above a negative threshold (PR24851)

Differential Revision: http://reviews.llvm.org/D14499

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252595 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAVX512 : Implemented encoding and DAG lowering for VMOVHPS/PD and VMOVLPS/PD instruct...
Igor Breger [Tue, 10 Nov 2015 07:09:07 +0000 (07:09 +0000)]
AVX512 : Implemented encoding and DAG lowering for VMOVHPS/PD and VMOVLPS/PD instructions.

Differential Revision: http://reviews.llvm.org/D14492

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252592 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove another variable unused in -Asserts build
David Blaikie [Tue, 10 Nov 2015 04:10:04 +0000 (04:10 +0000)]
Remove another variable unused in -Asserts build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252582 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove some unused variables to clean up the -Werror build
David Blaikie [Tue, 10 Nov 2015 03:16:28 +0000 (03:16 +0000)]
Remove some unused variables to clean up the -Werror build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252580 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding instruction aliases and tests.
Colin LeMahieu [Tue, 10 Nov 2015 01:58:26 +0000 (01:58 +0000)]
[Hexagon] Adding instruction aliases and tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252579 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSupport for emitting inline stack probes
Andy Ayers [Tue, 10 Nov 2015 01:50:49 +0000 (01:50 +0000)]
Support for emitting inline stack probes

For CoreCLR on Windows, stack probes must be emitted as inline sequences that probe successive stack pages
between the current stack limit and the desired new stack pointer location. This implements support for
the inline expansion on x64.

For in-body alloca probes, expansion is done during instruction lowering. For prolog probes, a stub call
is initially emitted during prolog creation, and expanded after epilog generation, to avoid complications
that arise when introducing new machine basic blocks during prolog and epilog creation.

Added a new test case, modified an existing one to exclude non-x64 coreclr (for now).

Add test case

Fix tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252578 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix mingw targets. Bandaid for r252532's buildbot brakage.
Richard Diamond [Tue, 10 Nov 2015 01:03:51 +0000 (01:03 +0000)]
Fix mingw targets. Bandaid for r252532's buildbot brakage.

Reviewers: brad.king, beanz

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14523

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252576 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Fixing compound register printing and reenabling more tests.
Colin LeMahieu [Tue, 10 Nov 2015 00:51:56 +0000 (00:51 +0000)]
[Hexagon] Fixing compound register printing and reenabling more tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252574 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAArch64: add experimental support for address tagging.
Tim Northover [Tue, 10 Nov 2015 00:44:23 +0000 (00:44 +0000)]
AArch64: add experimental support for address tagging.

AArch64 has the ability to use the top 8-bits of an "address" for extra
information, with the memory subsystem automatically masking them off for loads
and stores. When that's happening, we can sometimes skip masks on memory
operations in the compiler.

However, this requires the host OS and support stack to preserve those bits so
it can't be enabled everywhere. In principle iOS 8.0 and above do take the
required precautions and but we'll put it under a flag for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252573 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix llvm-nm(1) printing of llvm-bitcode files for -format darwin to match darwin...
Kevin Enderby [Tue, 10 Nov 2015 00:31:08 +0000 (00:31 +0000)]
Fix llvm-nm(1) printing of llvm-bitcode files for -format darwin to match darwin’s nm(1).
Also a small fix to match printing of Mach-O objects with -format posix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252567 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WebAssembly] Support 'unreachable' expression
Derek Schuff [Tue, 10 Nov 2015 00:30:57 +0000 (00:30 +0000)]
[WebAssembly] Support 'unreachable' expression

Lower LLVM's 'unreachable' terminator to ISD::TRAP, and lower ISD::TRAP to
wasm's 'unreachable' expression.

WebAssembly type-checks expressions, but a noreturn function with a
return type that doesn't match the context will cause a check
failure. So we lower LLVM 'unreachable' to ISD::TRAP and then lower that
to WebAssembly's 'unreachable' expression, which typechecks in any
context and causes a trap if executed.

Differential Revision: http://reviews.llvm.org/D14515

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252566 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove unnecessary call to getAllocatableRegClass
Matt Arsenault [Tue, 10 Nov 2015 00:30:14 +0000 (00:30 +0000)]
Remove unnecessary call to getAllocatableRegClass

I'm not sure what the point of this was. I'm not sure why
you would ever define an instruction that produces an unallocatable
register class. No tests fail with this removed, and it seems like
it should be a verifier error to define such an instruction.

This was problematic for AMDGPU because it would make bad decisions
by arbitrarily changing the register class when unsetting isAllocatable
for VS_32/VS_64, which is currently set as a workaround to this problem.

AMDGPU uses the VS_32/VS_64 register classes to represent operands which
can use either VGPRs or SGPRs. When  isAllocatable is unset for these,
this would need to pick  either the SGPR or VGPR class and insert either
a copy we don't want, or an illegal copy we would need to deal with
later. A semi-arbitrary register class ordering decision is made in tablegen,
which resulted in always picking a VGPR class because it happens to have
more registers than the SGPR register class. We really just want to
use whatever register class the original register had.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252565 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[PGO] Make indexed value profile data more compact
Xinliang David Li [Tue, 10 Nov 2015 00:24:45 +0000 (00:24 +0000)]
[PGO] Make indexed value profile data more compact

- Make indexed value profile data more compact by peeling out
  the per-site value count field into its own smaller sized array.
- Introduced formal data structure definitions to specify value
  profile data layout in indexed format. Previously the layout
  of the data is only assumed in the client code (scattered in
  three different places : size computation, EmitData, and ReadData
- The new data structure  serves as a central place for layout documentation.
- Add interfaces to force BE output for value profile data (testing purpose)
- Add byte swap unit tests

Differential Revision: http://reviews.llvm.org/D14401

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252563 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Fixing store instructions and reenabling a few more tests.
Colin LeMahieu [Tue, 10 Nov 2015 00:22:00 +0000 (00:22 +0000)]
[Hexagon] Fixing store instructions and reenabling a few more tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252561 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ARM] Handle t2ADDri in ARMAsmPrinter::EmitUnwindingInstruction.
Akira Hatanaka [Tue, 10 Nov 2015 00:10:41 +0000 (00:10 +0000)]
[ARM] Handle t2ADDri in ARMAsmPrinter::EmitUnwindingInstruction.

This fixes a bug in ARMAsmPrinter::EmitUnwindingInstruction where
llvm_unreachable was reached because t2ADDri wasn't handled.

Test case provided by Tim Northover.

rdar://problem/23270609

http://reviews.llvm.org/D14518

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252557 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Fixing load instruction parsing and reenabling tests.
Colin LeMahieu [Tue, 10 Nov 2015 00:02:27 +0000 (00:02 +0000)]
[Hexagon] Fixing load instruction parsing and reenabling tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252555 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMachineVerifier: Streamline live interval related error reporting
Matthias Braun [Mon, 9 Nov 2015 23:59:33 +0000 (23:59 +0000)]
MachineVerifier: Streamline live interval related error reporting

Simply perform additional report_context() calls after a report()
instead of adding more and more overloaded variations of report().  Also
improve several instances where information was output in an ad-hoc way
probably because no matching report() overload was available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252552 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMachineVerifier: Add missing linebreak
Matthias Braun [Mon, 9 Nov 2015 23:59:29 +0000 (23:59 +0000)]
MachineVerifier: Add missing linebreak

MachineInstr::print() with SkipOppers==true does not produce a
linebreak, so we have to do that in MachineVerifier::report().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252551 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMachineVerifier: MI::print has no TargetMachine overload
Matthias Braun [Mon, 9 Nov 2015 23:59:25 +0000 (23:59 +0000)]
MachineVerifier: MI::print has no TargetMachine overload

The code was passing a target machine pointer which degraded to a true
operand to SkipOppers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252550 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMachineVerifier: print list of live intervals if available
Matthias Braun [Mon, 9 Nov 2015 23:59:23 +0000 (23:59 +0000)]
MachineVerifier: print list of live intervals if available

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252549 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplify some APIs I was cleaning up while fixing -Wpessimizing-move warning
David Blaikie [Mon, 9 Nov 2015 23:51:45 +0000 (23:51 +0000)]
Simplify some APIs I was cleaning up while fixing -Wpessimizing-move warning

(Reid fixed the original error, but this seems nice to do in any case)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252548 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix -Wpessimizing-move warning in llvm-config.cpp
Reid Kleckner [Mon, 9 Nov 2015 23:37:26 +0000 (23:37 +0000)]
Fix -Wpessimizing-move warning in llvm-config.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252542 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WinEH] Remove isBarrier from instructions that do not return
Reid Kleckner [Mon, 9 Nov 2015 23:34:42 +0000 (23:34 +0000)]
[WinEH] Remove isBarrier from instructions that do not return

Fixes machine verification failures with David's latest EH change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252541 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoadd a SelectionDAG method to check if no common bits are set in two nodes; NFCI
Sanjay Patel [Mon, 9 Nov 2015 23:31:38 +0000 (23:31 +0000)]
add a SelectionDAG method to check if no common bits are set in two nodes; NFCI

This was suggested in:
http://reviews.llvm.org/D13956

and is a follow-on to:
http://reviews.llvm.org/rL252515
http://reviews.llvm.org/rL252519

This lets us remove logically equivalent/duplicated code from DAGCombiner and X86ISelDAGToDAG.

A corresponding function for IR instructions already exists in ValueTracking.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252539 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCodeGen: Be clear about semantics in SlotIndex::getNextSlot(), NFC
Duncan P. N. Exon Smith [Mon, 9 Nov 2015 23:31:01 +0000 (23:31 +0000)]
CodeGen: Be clear about semantics in SlotIndex::getNextSlot(), NFC

Be honest about using iterator semantics in `SlotIndex::getNextSlot()`
and `SlotIndex::getPrevSlot()`.  Instead of calling `getNextNode()` --
which is documented (but fails) to check for the sentinel -- call
`&*++getIterator()`.

This is (surprisingly!) a NFC commit.  `ilist_traits<IndexListEntry>`
has an `ilist_half_node<IndexListEntry>` as a sentinel (and no other
fields), and so the layout of `ilist<IndexListEntry>` is:
--
struct ilist<IndexListEntry> {
  ilist_half_node<IndexListEntry> Sentinel;
  IndexListEntry *Head;

  IndexListEntry *getHead() { return Head; }
  IndexListEntry *getSentinel() { return cast<...>(&Sentinel); }
};
--
In memory, this happens to look just like:
--
struct ilist<IndexListEntry> {
  ilist_node<IndexListEntry> Sentinel;

  IndexListEntry *getHead() { return Sentinel.getNext(); }
  IndexListEntry *getSentinel() { return cast<...>(&Sentinel); }
};
--
As a result, `ilist_node<IndexListEntry>::getNextNode()` that checks
`getNext()` of the possible sentinel will get a pointer to the head of
the list; it will never detect the sentinel, and will return the
sentinel itself instead of `nullptr` in the special cases.

Since `getNextNode()` and `getPrevNode()` don't work, just be honest
that we're not checking for the end/beginning of the list here.  Since
this code works, I guess we must never go past the sentinel.

(It's possible we're just getting lucky, and the new code will get
"lucky" in the same situations.  To properly fix that hypothetical bug,
we would need to check the iterator against `end()`/`begin()`.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252538 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix -Wdeprecated warnings due to the use of copy ops on SCEVPredicate derived class...
David Blaikie [Mon, 9 Nov 2015 23:30:15 +0000 (23:30 +0000)]
Fix -Wdeprecated warnings due to the use of copy ops on SCEVPredicate derived class objects

SCEVUnionPredicate is copied constructed here: lib/Transforms/Scalar/LoopDistribute.cpp:793
and move assigned (which can use the base class's copy ctor just
fine/without extra cost (I'd add it if it weren't for MSVC's issues
meaning = default is insufficient)) here: lib/Transforms/Utils/LoopVersioning.cpp:46

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252537 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[TargetLibraryInfo] Add support for fls, flsl, flsll.
Davide Italiano [Mon, 9 Nov 2015 23:23:20 +0000 (23:23 +0000)]
[TargetLibraryInfo] Add support for fls, flsl, flsll.

This is a prerequisite for further optimisations of these functions,
which will be commited as a separate patch.

Differential Revision: http://reviews.llvm.org/D14219

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252535 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[libFuzzer] make libFuzzer link if there is no sanitizer coverage instrumentation...
Kostya Serebryany [Mon, 9 Nov 2015 23:17:45 +0000 (23:17 +0000)]
[libFuzzer] make libFuzzer link if there is no sanitizer coverage instrumentation (it will fail at start-up time)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252533 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix `llvm-config` to adapt to the install environment.
Richard Diamond [Mon, 9 Nov 2015 23:15:38 +0000 (23:15 +0000)]
Fix `llvm-config` to adapt to the install environment.

Summary:
This patch does a couple of things:

  - Adds a new argument `--shared-mode` which accepts a list of components and prints whether or not the provided components need to be linked statically or shared.
  - Fixes `--libnames` when CMake BUILD_SHARED_LIBS is used.
  - Fixes `--libnames`, `--libs`, and `--libfiles` for dylib when static components aren't installed.
  - Fixes `--libnames`, `--libs`, `--libfiles`, and `--components` to use LLVM_DYLIB_COMPONENTS as the component manifest for dylib linking.
  - Uses the host platform's usual convention for filename extensions and such, instead of always defaulting to Unix-izms.

Because I don't own a Mac, I am not able to test the Mac platform dependent stuff locally. If someone would be willing to run a build for me on their machine (unless there's a better option), I'd appreciate it.

Reviewers: jfb, brad.king, whitequark, beanz

Subscribers: beanz, jauhien, llvm-commits

Differential Revision: http://reviews.llvm.org/D13198

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252532 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCombine ifdefs around dl_iterate_phdr in Unix/Signals.inc
Reid Kleckner [Mon, 9 Nov 2015 23:10:29 +0000 (23:10 +0000)]
Combine ifdefs around dl_iterate_phdr in Unix/Signals.inc

This avoids the need to have two dummy implementations of
findModulesAndOffsets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252531 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WinEH] Don't emit CATCHRET from visitCatchPad
David Majnemer [Mon, 9 Nov 2015 23:07:48 +0000 (23:07 +0000)]
[WinEH] Don't emit CATCHRET from visitCatchPad

Instead, emit a CATCHPAD node which will get selected to a target
specific sequence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252528 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDeprecate Autoconf
Chris Bieneman [Mon, 9 Nov 2015 21:54:55 +0000 (21:54 +0000)]
Deprecate Autoconf

As per the very positive feedback from llvm-dev (http://lists.llvm.org/pipermail/llvm-dev/2015-November/092150.html), this commit officially deprecates the LLVM autoconf-based build system.

Anyone still using it should switch to CMake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252520 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agospecify triple so Windows bots won't be sad
Sanjay Patel [Mon, 9 Nov 2015 21:53:58 +0000 (21:53 +0000)]
specify triple so Windows bots won't be sad

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252519 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd templated read/write to support::endian.
Rafael Espindola [Mon, 9 Nov 2015 21:34:45 +0000 (21:34 +0000)]
Add templated read/write to support::endian.

Expose read and write functions that take endianess as a template
parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252517 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] try harder to match bitwise 'or' into an LEA
Sanjay Patel [Mon, 9 Nov 2015 21:16:49 +0000 (21:16 +0000)]
[x86] try harder to match bitwise 'or' into an LEA

The motivation for this patch starts with the epic fail example in PR18007:
https://llvm.org/bugs/show_bug.cgi?id=18007

...unfortunately, this patch makes no difference for that case, but it solves some
simpler cases. We'll get there some day. :)

The current 'or' matching code was using computeKnownBits() via
isBaseWithConstantOffset() -> MaskedValueIsZero(), but that's an unnecessarily limited use.
We can do more by copying the logic in ValueTracking's haveNoCommonBitsSet(), so we can
treat the 'or' as if it was an 'add'.

There's a TODO comment here because we should lift the bit-checking logic into a helper
function, so it's not duplicated in DAGCombiner.

An example of the better LEA matching:

leal (%rdi,%rdi), %eax
andl $1, %esi
orl %esi, %eax

Becomes:

andl $1, %esi
leal (%rsi,%rdi,2), %eax

Differential Revision: http://reviews.llvm.org/D13956

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252515 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Separating statement to match what clang-format would do.
Colin LeMahieu [Mon, 9 Nov 2015 21:06:28 +0000 (21:06 +0000)]
[Hexagon] Separating statement to match what clang-format would do.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252513 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WinEH] Tweak funclet prologue/epilogue insertion to pass verifier
Reid Kleckner [Mon, 9 Nov 2015 21:04:00 +0000 (21:04 +0000)]
[WinEH] Tweak funclet prologue/epilogue insertion to pass verifier

For some reason we'd never run MachineVerifier on WinEH code, and you
explicitly have to ask for it with llc. I added it to a few test cases
to get some coverage.

Fixes PR25461.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252512 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[WinEH] Re-committing r252249 (Clone funclets with multiple parents) with additional...
Andrew Kaylor [Mon, 9 Nov 2015 19:59:02 +0000 (19:59 +0000)]
[WinEH] Re-committing r252249 (Clone funclets with multiple parents) with additional fixes for determinism problems

Differential Revision: http://reviews.llvm.org/D14454

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252508 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Fix -Wmicrosoft-enum-value warning with explicit enum type
Reid Kleckner [Mon, 9 Nov 2015 19:44:38 +0000 (19:44 +0000)]
[Hexagon] Fix -Wmicrosoft-enum-value warning with explicit enum type

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252505 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoWindows-specific test for sys::path::remove_dots.
Mike Aizatsky [Mon, 9 Nov 2015 19:36:53 +0000 (19:36 +0000)]
Windows-specific test for sys::path::remove_dots.

Differential Revision: http://reviews.llvm.org/D14503

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252504 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agodon't repeat function names in comments; NFC
Sanjay Patel [Mon, 9 Nov 2015 19:18:26 +0000 (19:18 +0000)]
don't repeat function names in comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252502 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agofix formatting; NFC
Sanjay Patel [Mon, 9 Nov 2015 18:59:30 +0000 (18:59 +0000)]
fix formatting; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252500 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMoving FileManager::removeDotPaths to llvm::sys::path::remove_dots
Mike Aizatsky [Mon, 9 Nov 2015 18:56:31 +0000 (18:56 +0000)]
Moving FileManager::removeDotPaths to llvm::sys::path::remove_dots

Differential Revision: http://reviews.llvm.org/D14393

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252499 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[sanitizer] Use same shadow offset for ASAN on aarch64
Adhemerval Zanella [Mon, 9 Nov 2015 18:03:48 +0000 (18:03 +0000)]
[sanitizer] Use same shadow offset for ASAN on aarch64

This patch makes ASAN for aarch64 use the same shadow offset for all
currently supported VMAs (39 and 42 bits).  The shadow offset is the
same for 39-bit (36).  Similar to ppc64 port, aarch64 transformation
also requires to use an add instead of 'or' for 42-bit VMA.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252495 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd discriminators for call instructions that are from the same line and same basic...
Dehao Chen [Mon, 9 Nov 2015 17:30:38 +0000 (17:30 +0000)]
Add discriminators for call instructions that are from the same line and same basic block.

Summary: Call instructions that are from the same line and same basic block needs to have separate discriminators to distinguish between different callsites.

Reviewers: davidxl, dnovillo, dblaikie

Subscribers: dblaikie, probinson, llvm-commits

Differential Revision: http://reviews.llvm.org/D14464

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252492 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplify. NFC.
Chad Rosier [Mon, 9 Nov 2015 16:56:06 +0000 (16:56 +0000)]
Simplify. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252491 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoGlobalOpt should maintain externally_initialized when splitting aggregates
Oliver Stannard [Mon, 9 Nov 2015 16:47:16 +0000 (16:47 +0000)]
GlobalOpt should maintain externally_initialized when splitting aggregates

When GlobalOpt splits an internal, global variable with an aggregate type, it
should propagate the externally_initialized flag to the newly created globals.

This makes the pass safe for our downstream use of this flag, while still
allowing some useful optimisations (such as removing dead parts of the split
aggregate) to be performed.

Differential Revision: http://reviews.llvm.org/D13382

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252490 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCreate Makefile variables for 'share' and 'libexec'
Jonathan Roelofs [Mon, 9 Nov 2015 16:10:00 +0000 (16:10 +0000)]
Create Makefile variables for 'share' and 'libexec'

in preparation for a Clang commit.

http://reviews.llvm.org/D14405

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252473 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LoopVectorize] Address post-commit feedback on r250032
James Molloy [Mon, 9 Nov 2015 14:32:05 +0000 (14:32 +0000)]
[LoopVectorize] Address post-commit feedback on r250032

Implemented as many of Michael's suggestions as were possible:
  * clang-format the added code while it is still fresh.
  * tried to change Value* to Instruction* in many places in computeMinimumValueSizes - unfortunately there are several places where Constants need to be handled so this wasn't possible.
  * Reduce the pass list on loop-vectorization-factors.ll.
  * Fix a bug where we were querying MinBWs for I->getOperand(0) but using MinBWs[I].

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252469 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAllow LLE/LD and the loop versioning infrastructure to use SCEV predicates
Silviu Baranga [Mon, 9 Nov 2015 13:26:09 +0000 (13:26 +0000)]
Allow LLE/LD and the loop versioning infrastructure to use SCEV predicates

Summary:
LAA currently generates a set of SCEV predicates that must be checked by users.
In the case of Loop Distribute/Loop Load Elimination, no such predicates could have
been emitted, since we don't allow stride versioning. However, in the future there
could be SCEV predicates that will need to be checked.

This change adds support for SCEV predicate versioning in the Loop Distribute, Loop
Load Eliminate and the loop versioning infrastructure.

Reviewers: anemet

Subscribers: mssimpso, sanjoy, llvm-commits

Differential Revision: http://reviews.llvm.org/D14240

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252467 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Add UABDL patterns for log2 shuffle.
Charlie Turner [Mon, 9 Nov 2015 13:10:52 +0000 (13:10 +0000)]
[AArch64] Add UABDL patterns for log2 shuffle.

Summary:
This matches the sum-of-absdiff patterns emitted by the vectoriser using log2 shuffles.

Relies on D14207 to be able to match the `extract_subvector(..., 0)`

Reviewers: t.p.northover, jmolloy

Subscribers: aemerson, llvm-commits, rengolin

Differential Revision: http://reviews.llvm.org/D14208

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252465 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Handle extract_subvector(..., 0) in ISel.
Charlie Turner [Mon, 9 Nov 2015 12:45:11 +0000 (12:45 +0000)]
[AArch64] Handle extract_subvector(..., 0) in ISel.

Summary:
Lowering this pattern early to an `EXTRACT_SUBREG` was making it impossible to match larger patterns in tblgen that use `extract_subvector(..., 0)` as part of the their input pattern.

It seems like there will exist somewhere a better way of specifying this pattern over all relevant register value types, but I didn't manage to find it.

Reviewers: t.p.northover, jmolloy

Subscribers: aemerson, llvm-commits, rengolin

Differential Revision: http://reviews.llvm.org/D14207

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252464 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[EABI] Add LLVM support for -meabi flag
Renato Golin [Mon, 9 Nov 2015 12:40:30 +0000 (12:40 +0000)]
[EABI] Add LLVM support for -meabi flag

"GCC requires the freestanding environment provide memcpy, memmove, memset
and memcmp": https://gcc.gnu.org/onlinedocs/gcc-5.2.0/gcc/Standards.html

Hence in GNUEABI targets LLVM should not convert 'memops' to their equivalent
'__aeabi_memops'. This convertion violates GCC contract.

The -meabi flag controls whether or not LLVM will modify 'memops' in GNUEABI
targets.

Without -meabi: use the triple default EABI.
With -meabi=default: use the triple default EABI.
With -meabi=gnu: use 'memops'.
With -meabi=4 or -meabi=5: use '__aeabi_memops'.
With -meabi set to an unknown value: same as -meabi=default.

Patch by Vinicius Tinti.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252462 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "[ARM] Combine CMOV into BFI where possible"
Renato Golin [Mon, 9 Nov 2015 12:19:10 +0000 (12:19 +0000)]
Revert "[ARM] Combine CMOV into BFI where possible"

This reverts commit r252057, as it broke ARM self-hosting buildbots, probably
due to a code-gen fault.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252460 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[CodeGen] Always promote f16 if not legal
Oliver Stannard [Mon, 9 Nov 2015 11:03:18 +0000 (11:03 +0000)]
[CodeGen] Always promote f16 if not legal

We don't currently have any runtime library functions for operations on
f16 values (other than conversions to and from f32 and f64), so we
should always promote it to f32, even if that is not a legal type. In
that case, the f32 values would be softened to f32 library calls.

SoftenFloatRes_FP_EXTEND now needs to check the promoted operand's type,
as it may ne a no-op or require a different library call.

getCopyFromParts and getCopyToParts now need to cope with a
floating-point value stored in a larger integer part, as is the case for
any target that needs to store an f16 value in a 32-bit integer
register.

Differential Revision: http://reviews.llvm.org/D12856

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252459 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding override to methods.
Colin LeMahieu [Mon, 9 Nov 2015 07:10:24 +0000 (07:10 +0000)]
[Hexagon] Adding override to methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252453 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Removing XFAIL on Hexagon target.
Colin LeMahieu [Mon, 9 Nov 2015 06:15:55 +0000 (06:15 +0000)]
[Hexagon] Removing XFAIL on Hexagon target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252450 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Fixing warnings.
Colin LeMahieu [Mon, 9 Nov 2015 05:47:56 +0000 (05:47 +0000)]
[Hexagon] Fixing warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252448 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Removing extra gen line.
Colin LeMahieu [Mon, 9 Nov 2015 05:31:39 +0000 (05:31 +0000)]
[Hexagon] Removing extra gen line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252447 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Maybe the makefile?
Colin LeMahieu [Mon, 9 Nov 2015 05:16:08 +0000 (05:16 +0000)]
[Hexagon] Maybe the makefile?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252446 91177308-0d34-0410-b5e6-96231b3b80d8