platform/upstream/llvm.git
8 years ago[Hexagon] Avoid register dependencies on indirect branches in packetizer
Krzysztof Parzyszek [Fri, 19 Aug 2016 21:07:35 +0000 (21:07 +0000)]
[Hexagon] Avoid register dependencies on indirect branches in packetizer

Do not packetize the instruction setting the branch address with the
indirect branch itself.

llvm-svn: 279324

8 years ago[ADT] add pointer_iterator, the opposite of pointee_iterator
Tim Shen [Fri, 19 Aug 2016 21:04:45 +0000 (21:04 +0000)]
[ADT] add pointer_iterator, the opposite of pointee_iterator

Differential Revision: https://reviews.llvm.org/D23703

llvm-svn: 279323

8 years agoRevert "[lldb][cmake] Remove libclang as an lldbBase dependency (NFCI)"
Vedant Kumar [Fri, 19 Aug 2016 21:00:40 +0000 (21:00 +0000)]
Revert "[lldb][cmake] Remove libclang as an lldbBase dependency (NFCI)"

This reverts commit r279296. Including LLDBDependencies breaks the
netbsd lldb bot because it exposes LLDB_USED_LIBS, which causes
lldb_link_common_libs to run to completion in unintended sites, which
results in a malformed call to target_link_libraries.

  http://lab.llvm.org:8011/builders/lldb-amd64-ninja-netbsd7/builds/5989

Thanks to Chris Bieneman for figuring this out!

llvm-svn: 279322

8 years ago[libFuzzer] fix the non-debug build warnings
Kostya Serebryany [Fri, 19 Aug 2016 20:57:09 +0000 (20:57 +0000)]
[libFuzzer] fix the non-debug build warnings

llvm-svn: 279321

8 years agoGlobalISel: translate floating-point round/extend
Tim Northover [Fri, 19 Aug 2016 20:48:23 +0000 (20:48 +0000)]
GlobalISel: translate floating-point round/extend

llvm-svn: 279320

8 years agoGlobalISel: translate floating-point comparisons
Tim Northover [Fri, 19 Aug 2016 20:48:16 +0000 (20:48 +0000)]
GlobalISel: translate floating-point comparisons

llvm-svn: 279319

8 years ago[asan] Temporarily mark test as broken on Windows
Vitaly Buka [Fri, 19 Aug 2016 20:48:09 +0000 (20:48 +0000)]
[asan] Temporarily mark test as broken on Windows

llvm-svn: 279318

8 years ago[NVPTX] Switch nvptx-use-infer-addrspace to true.
Justin Lebar [Fri, 19 Aug 2016 20:46:45 +0000 (20:46 +0000)]
[NVPTX] Switch nvptx-use-infer-addrspace to true.

Summary:
This switches us to use a different, more powerful algorithm for address
space inference.  I've tested this locally and it seems to work great.
Once we're more confident in it, we can remove the old pass altogether.

Reviewers: jingyue

Subscribers: llvm-commits, tra, jholewinski

Differential Revision: https://reviews.llvm.org/D23694

llvm-svn: 279317

8 years agoReverted premature changes to .clang-format
Kate Stone [Fri, 19 Aug 2016 20:46:37 +0000 (20:46 +0000)]
Reverted premature changes to .clang-format

llvm-svn: 279316

8 years agoMoved #include for lldb-python.h to a distinct group with a reminder comment
Kate Stone [Fri, 19 Aug 2016 20:44:07 +0000 (20:44 +0000)]
Moved #include for lldb-python.h to a distinct group with a reminder comment
declaring that it must be first.  Failure to do so results in build failures
on macOS due to subtle header conflicts.

llvm-svn: 279315

8 years agoReapply "ADT: Remove UB in ilist (and use a circular linked list)"
Duncan P. N. Exon Smith [Fri, 19 Aug 2016 20:40:12 +0000 (20:40 +0000)]
Reapply "ADT: Remove UB in ilist (and use a circular linked list)"

This reverts commit r279053, reapplying r278974 after fixing PR29035
with r279104.

Note that r279312 has been committed in the meantime, and this has been
rebased on top of that.  Otherwise it's identical to r278974.

Note for maintainers of out-of-tree code (that I missed in the original
message): if the new isKnownSentinel() assertion is firing from
ilist_iterator<>::operator*(), this patch has identified a bug in your
code.  There are a few common patterns:
- Some IR-related APIs htake an IRUnit* that might be nullptr, and pass
  in an incremented iterator as an insertion point.  Some old code was
  using "&*++I", which in the case of end() only worked by fluke.  If
  the IRUnit in question inherits from ilist_node_with_parent<>, you can
  use "I->getNextNode()".  Otherwise, use "List.getNextNode(*I)".
- In most other cases, crashes on &*I just need to check for I==end()
  before dereferencing.
- There's also occasional code that sends iterators into a function, and
  then starts calling I->getOperand() (or other API).  Either check for
  end() before the entering the function, or early exit.

Note for if the static_assert with HasObsoleteCustomization is firing
for you:
- r278513 has examples of how to stop using custom sentinel traits.
- r278532 removed ilist_nextprev_traits since no one was using it.  See
  lld's r278469 for the only migration I needed to do.

Original commit message follows.

----

This removes the undefined behaviour (UB) in ilist/ilist_node/etc.,
mainly by removing (gutting) the ilist_sentinel_traits customization
point and canonicalizing on a single, efficient memory layout.  This
fixes PR26753.

The new ilist is a doubly-linked circular list.
- ilist_node_base has two ilist_node_base*: Next and Prev.  Size-of: two
  pointers.
- ilist_node<T> (size-of: two pointers) is a type-safe wrapper around
  ilist_node_base.
- ilist_iterator<T> (size-of: two pointers) operates on an
  ilist_node<T>*, and downcasts to T* on dereference.
- ilist_sentinel<T> (size-of: two pointers) is a wrapper around
  ilist_node<T> that has some extra API for list management.
- ilist<T> (size-of: two pointers) has an ilist_sentinel<T>, whose
  address is returned for end().

The new memory layout matches ilist_half_embedded_sentinel_traits<T>
exactly.  The Head pointer that previously lived in ilist<T> is
effectively glued to the ilist_half_node<T> that lived in
ilist_half_embedded_sentinel_traits<T>, becoming the Next and Prev in
the ilist_sentinel_node<T>, respectively.  sizeof(ilist<T>) is now the
size of two pointers, and there is never any additional storage for a
sentinel.

This is a much simpler design for a doubly-linked list, removing most of
the corner cases of list manipulation (add, remove, etc.).  In follow-up
commits, I intend to move as many algorithms as possible into a
non-templated base class (ilist_base) to reduce code size.

Moreover, this fixes the UB in ilist_iterator/getNext/getPrev
operations.  Previously, ilist_iterator<T> operated on a T*, even when
the sentinel was not of type T (i.e., ilist_embedded_sentinel_traits and
ilist_half_embedded_sentinel_traits).  This added UB to all operations
involving end().   Now, ilist_iterator<T> operates on an ilist_node<T>*,
and only downcasts when the full type is guaranteed to be T*.

What did we lose?  There used to be a crash (in some configurations) on
++end().  Curiously (via UB), ++end() would return begin() for users of
ilist_half_embedded_sentinel_traits<T>, but otherwise ++end() would
cause a nice dependable nullptr dereference, crashing instead of a
possible infinite loop.  Options:
 1. Lose that behaviour.
 2. Keep it, by stealing a bit from Prev in asserts builds.
 3. Crash on dereference instead, using the same technique.

Hans convinced me (because of the number of problems this and r278532
exposed on Windows) that we really need some assertion here, at least in
the short term.  I've opted for #3 since I think it catches more bugs.

I added only a couple of unit tests to root out specific bugs I hit
during bring-up, but otherwise this is tested implicitly via the
extensive usage throughout LLVM.

Planned follow-ups:
- Remove ilist_*sentinel_traits<T>.  Here I've just gutted them to
  prevent build failures in sub-projects.  Once I stop referring to them
  in sub-projects, I'll come back and delete them.
- Add ilist_base and move algorithms there.
- Check and fix move construction and assignment.

Eventually, there are other interesting directions:
- Rewrite reverse iterators, so that rbegin().getNodePtr()==&*rbegin().
  This allows much simpler logic when erasing elements during a reverse
  traversal.
- Remove ilist_traits::createNode, by deleting the remaining API that
  creates nodes.  Intrusive lists shouldn't be creating nodes
  themselves.
- Remove ilist_traits::deleteNode, by (1) asserting that lists are empty
  on destruction and (2) changing API that calls it to take a Deleter
  functor (intrusive lists shouldn't be in the memory management
  business).
- Reconfigure the remaining callback traits (addNodeToList, etc.) to be
  higher-level, pulling out a simple_ilist<T> that is much easier to
  read and understand.
- Allow tags (e.g., ilist_node<T,tag1> and ilist_node<T,tag2>) so that T
  can be a member of multiple intrusive lists.

llvm-svn: 279314

8 years agoRevert "[SimplifyCFG] Rewrite SinkThenElseCodeToEnd"
Reid Kleckner [Fri, 19 Aug 2016 20:22:39 +0000 (20:22 +0000)]
Revert "[SimplifyCFG] Rewrite SinkThenElseCodeToEnd"

This reverts commit r279229. It breaks intrinsic function calls in
diamonds.

llvm-svn: 279313

8 years agoReapply "ADT: Tidy up ilist_traits static asserts, NFC"
Duncan P. N. Exon Smith [Fri, 19 Aug 2016 20:17:23 +0000 (20:17 +0000)]
Reapply "ADT: Tidy up ilist_traits static asserts, NFC"

This spiritually reapplies r279012 (reverted in r279052) without the
r278974 parts.  The differences:

  - Only the HasGetNext trait exists here, so I've only cleaned up (and
    tested) it.  I still added HasObsoleteCustomization since I know
    this will be expanding when r278974 is reapplied.

  - I changed the unit tests to use static_assert to catch problems
    earlier in the build.

  - I added negative tests for the type traits.

Original commit message follows.

----

Change the ilist traits to use decltype instead of sizeof, and add
HasObsoleteCustomization so that additions to this list don't
need to be added in two places.

I suspect this will now work with MSVC, since the trait tested in
r278991 seems to work.  If for some reason it continues to fail on
Windows I'll follow up by adding back the #ifndef _MSC_VER.

llvm-svn: 279312

8 years agoGlobalISel: translate floating-point constants
Tim Northover [Fri, 19 Aug 2016 20:09:15 +0000 (20:09 +0000)]
GlobalISel: translate floating-point constants

llvm-svn: 279311

8 years agoGlobalISel: translate float/int conversion instructions.
Tim Northover [Fri, 19 Aug 2016 20:09:11 +0000 (20:09 +0000)]
GlobalISel: translate float/int conversion instructions.

llvm-svn: 279310

8 years agoGlobalISel: support translating select instructions.
Tim Northover [Fri, 19 Aug 2016 20:09:07 +0000 (20:09 +0000)]
GlobalISel: support translating select instructions.

llvm-svn: 279309

8 years agoGlobalISel: fix insert/extract to work on ConstantExprs too.
Tim Northover [Fri, 19 Aug 2016 20:09:03 +0000 (20:09 +0000)]
GlobalISel: fix insert/extract to work on ConstantExprs too.

No tests yet unfortunately (ConstantFolding reduces all supported constants to
ConstantInts before we get to translation). Soon.

llvm-svn: 279308

8 years agoGlobalISel: fix stale comment
Tim Northover [Fri, 19 Aug 2016 20:09:01 +0000 (20:09 +0000)]
GlobalISel: fix stale comment

llvm-svn: 279307

8 years agoGlobalISel: translate insertvalue instructions.
Tim Northover [Fri, 19 Aug 2016 20:08:55 +0000 (20:08 +0000)]
GlobalISel: translate insertvalue instructions.

This adds a G_INSERT instruction, which technically makes G_SEQUENCE redundant
(it's equivalent to a G_INSERT into an IMPLICIT_DEF). We'll leave G_SEQUENCE
for now though: it's likely to be far more common as it's a fundamental part of
legalization, so avoiding the mess and bloat of the extra IMPLICIT_DEFs is
probably worthwhile.

llvm-svn: 279306

8 years agoMachineScheduler: Add constructor functions for the DAGMutations
Tom Stellard [Fri, 19 Aug 2016 19:59:18 +0000 (19:59 +0000)]
MachineScheduler: Add constructor functions for the DAGMutations

Summary: This way they can be re-used by target-specific schedulers.

Reviewers: atrick, MatzeB, kparzysz

Subscribers: kparzysz, llvm-commits, MatzeB

Differential Revision: https://reviews.llvm.org/D23678

llvm-svn: 279305

8 years ago[Hexagon] Add RUN line to test
Krzysztof Parzyszek [Fri, 19 Aug 2016 19:36:35 +0000 (19:36 +0000)]
[Hexagon] Add RUN line to test

llvm-svn: 279304

8 years ago[Hexagon] Fix subesthetic indentation
Krzysztof Parzyszek [Fri, 19 Aug 2016 19:29:15 +0000 (19:29 +0000)]
[Hexagon] Fix subesthetic indentation

llvm-svn: 279303

8 years ago[Hexagon] Allow i1 values for 'r' constraint in inline-asm
Krzysztof Parzyszek [Fri, 19 Aug 2016 19:17:28 +0000 (19:17 +0000)]
[Hexagon] Allow i1 values for 'r' constraint in inline-asm

llvm-svn: 279302

8 years ago[CostModel][X86] Added sub, or, and, fadd and fsub costs and missing 512-bit mul...
Simon Pilgrim [Fri, 19 Aug 2016 19:07:10 +0000 (19:07 +0000)]
[CostModel][X86] Added sub, or, and, fadd and fsub costs and missing 512-bit mul costs

llvm-svn: 279301

8 years ago[InstCombine] remove an icmp fold that is already handled by InstSimplify
Sanjay Patel [Fri, 19 Aug 2016 19:03:07 +0000 (19:03 +0000)]
[InstCombine] remove an icmp fold that is already handled by InstSimplify

Specifically, this is done near the end of "SimplifyICmpInst" using
computeKnownBits() as the broader solution. There are even vector
tests (yay!) for this in test/Transforms/InstSimplify/compare.ll.

I considered putting an assert here instead of just deleting, but
then we could assert every possible fold in InstSimplify in
InstCombine, so...less is more?

llvm-svn: 279300

8 years agoAdd missing #include found by modules build.
Richard Smith [Fri, 19 Aug 2016 18:57:17 +0000 (18:57 +0000)]
Add missing #include found by modules build.

llvm-svn: 279298

8 years ago[Hexagon] Do not cache alloca instructions during isel
Krzysztof Parzyszek [Fri, 19 Aug 2016 18:46:13 +0000 (18:46 +0000)]
[Hexagon] Do not cache alloca instructions during isel

They can be deleted or replicated, so the cache may become outdated.
They only need to be visited once during frame lowering, so just scan
the function instead.

llvm-svn: 279297

8 years ago[lldb][cmake] Remove libclang as an lldbBase dependency (NFCI)
Vedant Kumar [Fri, 19 Aug 2016 18:41:50 +0000 (18:41 +0000)]
[lldb][cmake] Remove libclang as an lldbBase dependency (NFCI)

It's pulling in all kinds of things it doesn't need (e.g, clang-tidy!).

Eliminating this dependency removes 1056 dependencies from the
'CommandObjectFrame.cpp.o' target and 454 dependencies from the 'lldb'
target. On my machine, this shaves 7 minutes off of a clean build of
lldb.

Thanks to Zachary Turner for pointing out some issues with an earlier
version of this patch!

Differential Revision: https://reviews.llvm.org/D22987

llvm-svn: 279296

8 years ago[PM] Re-instate r279227 and r279228 with a fix to the way the templating
Chandler Carruth [Fri, 19 Aug 2016 18:36:06 +0000 (18:36 +0000)]
[PM] Re-instate r279227 and r279228 with a fix to the way the templating
was done to hopefully appease MSVC.

As an upside, this also implements the suggestion Sanjoy made in code
review, so two for one! =]

I'll be watching the bots to see if there are still issues.

llvm-svn: 279295

8 years agoGlobalISel: improve representation of G_SEQUENCE and G_EXTRACT
Tim Northover [Fri, 19 Aug 2016 18:32:14 +0000 (18:32 +0000)]
GlobalISel: improve representation of G_SEQUENCE and G_EXTRACT

First, make sure all types involved are represented, rather than being implicit
from the register width.

Second, canonicalize all types to scalar. These operations just act in bits and
don't care about vectors.

Also standardize spelling of Indices in the MachineIRBuilder (NFC here).

llvm-svn: 279294

8 years ago[CostModel][X86] Added some AVX512 and 512-bit vector cost tests
Simon Pilgrim [Fri, 19 Aug 2016 18:24:10 +0000 (18:24 +0000)]
[CostModel][X86] Added some AVX512 and 512-bit vector cost tests

llvm-svn: 279291

8 years agoRevert "IfConversion: Rescan diamonds."
Kyle Butt [Fri, 19 Aug 2016 18:17:06 +0000 (18:17 +0000)]
Revert "IfConversion: Rescan diamonds."

This reverts commit bfd62a4b4465dd21811bf615c3b04c30ddb09f7b.

llvm-svn: 279289

8 years agoRevert "CodeGen: If Convert blocks that would form a diamond when tail-merged."
Kyle Butt [Fri, 19 Aug 2016 18:17:04 +0000 (18:17 +0000)]
Revert "CodeGen: If Convert blocks that would form a diamond when tail-merged."

This reverts commit 0fda93481c4231c06b838ef476c0c404c51ff875.

llvm-svn: 279288

8 years agoGlobalISel: allow extractvalue to extract an aggregate.
Tim Northover [Fri, 19 Aug 2016 18:09:41 +0000 (18:09 +0000)]
GlobalISel: allow extractvalue to extract an aggregate.

llvm-svn: 279287

8 years ago[Hexagon] Fixes for new-value jump formation
Krzysztof Parzyszek [Fri, 19 Aug 2016 17:54:49 +0000 (17:54 +0000)]
[Hexagon] Fixes for new-value jump formation

- Recognize C2_cmpgtui, S2_tstbit_i, and S4_ntstbit_i.
- Avoid creating new-value instructions with both source operands equal.

llvm-svn: 279286

8 years agoGlobalISel: support translation of extractvalue instructions.
Tim Northover [Fri, 19 Aug 2016 17:47:05 +0000 (17:47 +0000)]
GlobalISel: support translation of extractvalue instructions.

llvm-svn: 279285

8 years ago[CostModel][X86] Add fdiv + frem cost tests
Simon Pilgrim [Fri, 19 Aug 2016 17:39:00 +0000 (17:39 +0000)]
[CostModel][X86] Add fdiv + frem cost tests

llvm-svn: 279283

8 years ago[InstCombine] use local variables to reduce code in foldICmpShlConstant; NFC
Sanjay Patel [Fri, 19 Aug 2016 17:34:05 +0000 (17:34 +0000)]
[InstCombine] use local variables to reduce code in foldICmpShlConstant; NFC

llvm-svn: 279282

8 years ago[Hexagon] Fix a few omissions in HexagonInstrInfo
Krzysztof Parzyszek [Fri, 19 Aug 2016 17:20:57 +0000 (17:20 +0000)]
[Hexagon] Fix a few omissions in HexagonInstrInfo

llvm-svn: 279280

8 years ago[InstCombine] rename variables in foldICmpShlConstant(); NFC
Sanjay Patel [Fri, 19 Aug 2016 17:20:37 +0000 (17:20 +0000)]
[InstCombine] rename variables in foldICmpShlConstant(); NFC

llvm-svn: 279279

8 years agoGlobalISel: support overflow arithmetic intrinsics.
Tim Northover [Fri, 19 Aug 2016 17:17:06 +0000 (17:17 +0000)]
GlobalISel: support overflow arithmetic intrinsics.

Unsigned addition and subtraction can reuse the instructions created to
legalize large width operations (i.e. both produce and consume a carry flag).
Signed operations and multiplies get a dedicated op-with-overflow instruction.

Once this is produced the two values are combined into a struct register (which
will almost always be merged with a corresponding G_EXTRACT as part of
legalization).

llvm-svn: 279278

8 years agoRevert "[asan] Optimize store size in FunctionStackPoisoner::poisonRedZones"
Vitaly Buka [Fri, 19 Aug 2016 17:15:38 +0000 (17:15 +0000)]
Revert "[asan] Optimize store size in FunctionStackPoisoner::poisonRedZones"

This reverts commit r279178.

Speculative revert in hope to fix asan crash on arm.

llvm-svn: 279277

8 years agoRevert "[asan] Fix size of shadow incorrectly calculated in r279178"
Vitaly Buka [Fri, 19 Aug 2016 17:15:33 +0000 (17:15 +0000)]
Revert "[asan] Fix size of shadow incorrectly calculated in r279178"

This reverts commit r279222.

Speculative revert in hope to fix asan crash on arm.

llvm-svn: 279276

8 years ago[RuntimeDyld] Revert r279182 and 279201 -- they broke some ARM bots.
Lang Hames [Fri, 19 Aug 2016 17:06:39 +0000 (17:06 +0000)]
[RuntimeDyld] Revert r279182 and 279201 -- they broke some ARM bots.

llvm-svn: 279275

8 years ago[AliasSetTracker] Degrade AliasSetTracker when may-alias sets get too large.
Michael Kuperstein [Fri, 19 Aug 2016 17:05:22 +0000 (17:05 +0000)]
[AliasSetTracker] Degrade AliasSetTracker when may-alias sets get too large.

Repeated inserts into AliasSetTracker have quadratic behavior - inserting a
pointer into AST is linear, since it requires walking over all "may" alias
sets and running an alias check vs. every pointer in the set.

We can avoid this by tracking the total number of pointers in "may" sets,
and when that number exceeds a threshold, declare the tracker "saturated".
This lumps all pointers into a single "may" set that aliases every other
pointer.

(This is a stop-gap solution until we migrate to MemorySSA)

This fixes PR28832.
Differential Revision: https://reviews.llvm.org/D23432

llvm-svn: 279274

8 years ago[X86][SSE] Generalised combining to VZEXT_MOVL to any vector size
Simon Pilgrim [Fri, 19 Aug 2016 17:02:00 +0000 (17:02 +0000)]
[X86][SSE] Generalised combining to VZEXT_MOVL to any vector size

This doesn't change tests codegen as we already combined to blend+zero which is what we lower VZEXT_MOVL to on SSE41+ targets, but it does put us in a better position when we improve shuffling for optsize.

llvm-svn: 279273

8 years ago[Hexagon] Enforce LLSC packetization rules
Krzysztof Parzyszek [Fri, 19 Aug 2016 16:57:05 +0000 (16:57 +0000)]
[Hexagon] Enforce LLSC packetization rules

Ensure that load locked and store conditional instructions are only
packetized with ALU32 instructions.

Patch by Ben Craig.

llvm-svn: 279272

8 years agoFix regression in InstCombine introduced by r278944
Reid Kleckner [Fri, 19 Aug 2016 16:53:18 +0000 (16:53 +0000)]
Fix regression in InstCombine introduced by r278944

The intended transform is:
  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
  // -> and (icmp eq P, null), (icmp eq Q, null).

P and Q are both pointer types, but may have different types. We need
two calls to getNullValue() to make the icmps.

llvm-svn: 279271

8 years agoMachineScheduler: Make some GenericScheduler member variables protected
Tom Stellard [Fri, 19 Aug 2016 16:44:32 +0000 (16:44 +0000)]
MachineScheduler: Make some GenericScheduler member variables protected

Summary: We will need these in AMDGPU's new SchedStrategy implmentation.

Reviewers: MatzeB, atrick

Subscribers: llvm-commits, MatzeB

Differential Revision: https://reviews.llvm.org/D23679

llvm-svn: 279270

8 years ago[Hexagon] Minor updates to register definitions
Krzysztof Parzyszek [Fri, 19 Aug 2016 16:40:19 +0000 (16:40 +0000)]
[Hexagon] Minor updates to register definitions

llvm-svn: 279269

8 years ago[CloneFunction] Don't remove unrelated nodes from the CGSSC
David Majnemer [Fri, 19 Aug 2016 16:37:40 +0000 (16:37 +0000)]
[CloneFunction] Don't remove unrelated nodes from the CGSSC

CGSCC use a WeakVH to track call sites.  RAUW a call within a function
can result in that WeakVH getting confused about whether or not the call
site is still around.

llvm-svn: 279268

8 years ago[Hexagon] Fix incorrect generation of S4_subi_asl_ri
Krzysztof Parzyszek [Fri, 19 Aug 2016 16:35:05 +0000 (16:35 +0000)]
[Hexagon] Fix incorrect generation of S4_subi_asl_ri

Patch by Jyotsna Verma.

llvm-svn: 279267

8 years ago[InstCombine] use m_APInt to allow icmp (shl 1, Y), C folds for splat constant vectors
Sanjay Patel [Fri, 19 Aug 2016 16:12:16 +0000 (16:12 +0000)]
[InstCombine] use m_APInt to allow icmp (shl 1, Y), C folds for splat constant vectors

llvm-svn: 279266

8 years ago[Hexagon] Add missing pattern for C4_cmplte
Krzysztof Parzyszek [Fri, 19 Aug 2016 16:11:33 +0000 (16:11 +0000)]
[Hexagon] Add missing pattern for C4_cmplte

llvm-svn: 279265

8 years ago[ELF] - Fix for PR28976 - Corrupted section contents when using linker scripts
George Rimar [Fri, 19 Aug 2016 15:46:28 +0000 (15:46 +0000)]
[ELF] - Fix for PR28976 - Corrupted section contents when using linker scripts

This is fix for PR28976.

Problem was that in scanRelocs, we computed relocation offset too early
for case when linkerscript was used. Patch fixes the issue
delaying the calculation.

Differential revision: https://reviews.llvm.org/D23655

llvm-svn: 279264

8 years ago[InstCombine] use m_APInt to allow icmp X, C folds for splat constant vectors
Sanjay Patel [Fri, 19 Aug 2016 15:40:44 +0000 (15:40 +0000)]
[InstCombine] use m_APInt to allow icmp X, C folds for splat constant vectors

Of course, we really need to refactor and fix all of the cmp predicates,
but this one is interesting because without it, we later perform an
information-losing transform of icmp (shl 1, Y), C, and we can't recover
the better fold.

llvm-svn: 279263

8 years ago[ELF] - Give automatically generated __start_* and __stop_* symbols default visibility.
George Rimar [Fri, 19 Aug 2016 15:36:32 +0000 (15:36 +0000)]
[ELF] - Give automatically generated __start_* and __stop_* symbols default visibility.

This patch is opposite to D19024, which made this symbols to be hidden by default.

Unfortunately FreeBSD loader wants to see
start_set_modmetadata_set/stop_set_modmetadata_set in the dynamic symbol table.
They were not placed there because had hidden visibility.

Patch makes them to have default visibility again.

Differential revision: https://reviews.llvm.org/D23552

llvm-svn: 279262

8 years ago[LTO] Remove dead-code: collectUsedGlobalVariables has been moved to Thin and LTO...
Mehdi Amini [Fri, 19 Aug 2016 15:35:44 +0000 (15:35 +0000)]
[LTO] Remove dead-code: collectUsedGlobalVariables has been moved to Thin and LTO specifc path (NFC)

llvm-svn: 279261

8 years ago[InstCombine] add tests for missing vector icmp folds
Sanjay Patel [Fri, 19 Aug 2016 15:27:28 +0000 (15:27 +0000)]
[InstCombine] add tests for missing vector icmp folds

llvm-svn: 279259

8 years ago[ELF] - Linkerscript: set correct dynamic tag entries values when LS is used.
George Rimar [Fri, 19 Aug 2016 15:23:39 +0000 (15:23 +0000)]
[ELF] - Linkerscript: set correct dynamic tag entries values when LS is used.

Previously DT_PREINIT_ARRAYSZ, DT_INIT_ARRAYSZ and DT_FINI_ARRAYSZ
were set to zero when lincerscript was used becase sections sizes are
calculated later that were taken.

Patch delays values calculation for these entries. Testcase is provided.

Differential revision: https://reviews.llvm.org/D23663

llvm-svn: 279258

8 years ago[InstCombine] add missing tests for basic icmp folds
Sanjay Patel [Fri, 19 Aug 2016 15:21:45 +0000 (15:21 +0000)]
[InstCombine] add missing tests for basic icmp folds

These are implicitly included as part of larger test cases, but they don't
exist stand-alone (and don't happen for vectors...).

llvm-svn: 279257

8 years ago[ELF] - Linkerscript: implemented SUBALIGN() command.
George Rimar [Fri, 19 Aug 2016 15:18:23 +0000 (15:18 +0000)]
[ELF] - Linkerscript: implemented SUBALIGN() command.

You can force input section alignment within an output section by using SUBALIGN. The
value specified overrides any alignment given by input sections, whether larger or smaller.

SUBALIGN is used in many projects in the wild.

Differential revision: https://reviews.llvm.org/D23063

llvm-svn: 279256

8 years ago[Hexagon] Make p0 an explicit operand in VA1_clr* subinstructions, NFC
Krzysztof Parzyszek [Fri, 19 Aug 2016 15:17:19 +0000 (15:17 +0000)]
[Hexagon] Make p0 an explicit operand in VA1_clr* subinstructions, NFC

llvm-svn: 279255

8 years ago[Hexagon] Add explicit default constructor for HexagonSelectionDAGInfo
Krzysztof Parzyszek [Fri, 19 Aug 2016 15:13:54 +0000 (15:13 +0000)]
[Hexagon] Add explicit default constructor for HexagonSelectionDAGInfo

llvm-svn: 279254

8 years agobuild: allow building a specific set of sanitizers
Saleem Abdulrasool [Fri, 19 Aug 2016 15:13:21 +0000 (15:13 +0000)]
build: allow building a specific set of sanitizers

Introduce a new CMake option `COMPILER_RT_SANITIZERS_TO_BUILD` which takes
either a special token `all` (default) which will preserve the current behaviour
or a CMake list of sanitizers to build.  It will still perform the normal checks
if the sanitizer is requested.  It only permits a further means to exclude a
particular sanitizer.  This gives finer grained control than
`COMPILER_RT_BUILD_SANITIZERS` which only gives an all or nothing control.

llvm-svn: 279253

8 years agoUnxfail passing tests on Hexagon
Krzysztof Parzyszek [Fri, 19 Aug 2016 15:07:58 +0000 (15:07 +0000)]
Unxfail passing tests on Hexagon

llvm-svn: 279252

8 years ago[Hexagon] Allow tail-call optimization when mixing C and fast calling conv
Krzysztof Parzyszek [Fri, 19 Aug 2016 15:02:18 +0000 (15:02 +0000)]
[Hexagon] Allow tail-call optimization when mixing C and fast calling conv

Patch by Arnold Schwaighofer.

llvm-svn: 279251

8 years agoclang-format: [JS] handle object literals with casts.
Martin Probst [Fri, 19 Aug 2016 14:35:01 +0000 (14:35 +0000)]
clang-format: [JS] handle object literals with casts.

Summary: E.g. `{a: 1} as b`.

Reviewers: djasper

Subscribers: cfe-commits, klimek

Differential Revision: https://reviews.llvm.org/D23714

llvm-svn: 279250

8 years ago[Hexagon] Check for empty live interval
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:29:43 +0000 (14:29 +0000)]
[Hexagon] Check for empty live interval

Patch by Brendon Cahoon.

llvm-svn: 279249

8 years ago[Hexagon] Consider zext/sext of a load to i32 to be free
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:22:07 +0000 (14:22 +0000)]
[Hexagon] Consider zext/sext of a load to i32 to be free

llvm-svn: 279248

8 years agoRevert r279242 - it's failing the tests
Anton Korobeynikov [Fri, 19 Aug 2016 14:18:34 +0000 (14:18 +0000)]
Revert r279242 - it's failing the tests

llvm-svn: 279247

8 years ago[Hexagon] Handle J2_jumptpt and J2_jumpfpt instructions
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:14:09 +0000 (14:14 +0000)]
[Hexagon] Handle J2_jumptpt and J2_jumpfpt instructions

llvm-svn: 279246

8 years ago[Hexagon] Fix indentation, NFC
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:12:51 +0000 (14:12 +0000)]
[Hexagon] Fix indentation, NFC

llvm-svn: 279245

8 years ago[Hexagon] Remove unnecessary llvm::, NFC
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:10:57 +0000 (14:10 +0000)]
[Hexagon] Remove unnecessary llvm::, NFC

llvm-svn: 279244

8 years ago[Hexagon] Rename the HEXAGON_MC namespace to Hexagon_MC, NFC
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:09:47 +0000 (14:09 +0000)]
[Hexagon] Rename the HEXAGON_MC namespace to Hexagon_MC, NFC

llvm-svn: 279243

8 years agoFix PR27500: on MSP430 the branch destination offset is measured in words, not bytes.
Anton Korobeynikov [Fri, 19 Aug 2016 14:07:10 +0000 (14:07 +0000)]
Fix PR27500: on MSP430 the branch destination offset is measured in words, not bytes.
In addition, the branch instructions will have proper BB destinations, not offsets, like before.

Patch by Vadzim Dambrouski!

Differential Revision: https://reviews.llvm.org/D20162

llvm-svn: 279242

8 years ago[Hexagon] Mark PS_jumpret as pseudo-instruction, expand it into J2_jumpr
Krzysztof Parzyszek [Fri, 19 Aug 2016 14:04:45 +0000 (14:04 +0000)]
[Hexagon] Mark PS_jumpret as pseudo-instruction, expand it into J2_jumpr

llvm-svn: 279241

8 years agoFixed a typo (compilation_commands.json --> compile_commands.json).
Andrey Bokhanko [Fri, 19 Aug 2016 13:36:31 +0000 (13:36 +0000)]
Fixed a typo (compilation_commands.json --> compile_commands.json).

llvm-svn: 279240

8 years ago[Hexagon] Improvements to handling and generation of FP instructions
Krzysztof Parzyszek [Fri, 19 Aug 2016 13:34:31 +0000 (13:34 +0000)]
[Hexagon] Improvements to handling and generation of FP instructions

Improved handling of fma, floating point min/max, additional load/store
instructions for floating point types.

Patch by Jyotsna Verma.

llvm-svn: 279239

8 years agoFix 32-bit builds after r279232
Pavel Labath [Fri, 19 Aug 2016 13:14:13 +0000 (13:14 +0000)]
Fix 32-bit builds after r279232

GetByteSize() of a DataBuffer returns a uint64_t (it probably shouldn't), which isn't implicitly
convertible to size_t.

llvm-svn: 279238

8 years agoCleanup: Move the *AddressDescription printing code to Print() members inside those...
Filipe Cabecinhas [Fri, 19 Aug 2016 13:07:23 +0000 (13:07 +0000)]
Cleanup: Move the *AddressDescription printing code to Print() members inside those structs.

Summary:
The Print() members might take optional access_size and bug_type
parameters to still be able to provide the same information

Reviewers: kcc, samsonov

Subscribers: kubabrecka, llvm-commits

Differential Revision: https://reviews.llvm.org/D23658

llvm-svn: 279237

8 years ago[ELF] - Removed excessive llvm-readobj call from linkerscript-ouputformat.s. NFC.
George Rimar [Fri, 19 Aug 2016 12:57:45 +0000 (12:57 +0000)]
[ELF] - Removed excessive llvm-readobj call from linkerscript-ouputformat.s. NFC.

llvm-svn: 279236

8 years ago[AMDGPU] add s_incperflevel/s_decperflevel builtins
Valery Pykhtin [Fri, 19 Aug 2016 12:54:31 +0000 (12:54 +0000)]
[AMDGPU] add s_incperflevel/s_decperflevel builtins

Differential revision: https://reviews.llvm.org/D23668

llvm-svn: 279235

8 years agoFixing a Darwing test thats failing on windows
Dimitar Vlahovski [Fri, 19 Aug 2016 12:44:53 +0000 (12:44 +0000)]
Fixing a Darwing test thats failing on windows

The pexpect import should be make after the skip-if-not-darwin part
because pexpect is not available on Windows

llvm-svn: 279234

8 years ago[LoopVectorize] Don't copy std::vector in for-range loop.
Benjamin Kramer [Fri, 19 Aug 2016 12:44:24 +0000 (12:44 +0000)]
[LoopVectorize] Don't copy std::vector in for-range loop.

llvm-svn: 279233

8 years agoRemove the last manually constructed packet from gdb-remote register context + small...
Pavel Labath [Fri, 19 Aug 2016 12:31:49 +0000 (12:31 +0000)]
Remove the last manually constructed packet from gdb-remote register context + small refactor

Summary:
The tricky part here was that the exisiting implementation of WriteAllRegisters was expecting
hex-encoded data (as that was what the first implementation I replaced was using, but here we had
binary data to begin with. I thought the read/write register functions would be more useful if
they handled the hex-encoding themselves (all the other client functions provide the responses in
a more-or-less digested form). The read functions return a DataBuffer, so they can allocate as
much memory as they need to, while the write functions functions take an llvm::ArrayRef, as that
can be constructed from pretty much anything.

Reviewers: clayborg

Subscribers: lldb-commits

Differential Revision: https://reviews.llvm.org/D23659

llvm-svn: 279232

8 years ago[PM] Revert r279227 and r279228 until I can find someone to help me
Chandler Carruth [Fri, 19 Aug 2016 10:51:55 +0000 (10:51 +0000)]
[PM] Revert r279227 and r279228 until I can find someone to help me
solve completely opaque MSVC build errors. It complains about lots of
stuff with this change without givin nearly enough information to even
try to fix.

llvm-svn: 279231

8 years ago[X86][SSE] Add support for matching commuted insertps patterns
Simon Pilgrim [Fri, 19 Aug 2016 10:31:53 +0000 (10:31 +0000)]
[X86][SSE] Add support for matching commuted insertps patterns

INSERTPS doesn't fit well with our shuffle mask canonicalization, so we need to attempt both the original mask and the commuted mask to more likely get a match

llvm-svn: 279230

8 years ago[SimplifyCFG] Rewrite SinkThenElseCodeToEnd
James Molloy [Fri, 19 Aug 2016 10:10:27 +0000 (10:10 +0000)]
[SimplifyCFG] Rewrite SinkThenElseCodeToEnd

The new version has several advantages:
  1) IMSHO it's more readable and neater
  2) It handles loads and stores properly
  3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.

With this change we can now finally sink load-modify-store idioms such as:

    if (a)
      return *b += 3;
    else
      return *b += 4;

    =>

    %z = load i32, i32* %y
    %.sink = select i1 %a, i32 5, i32 7
    %b = add i32 %z, %.sink
    store i32 %b, i32* %y
    ret i32 %b

When this works for switches it'll be even more powerful.

llvm-svn: 279229

8 years ago[PM] Fix a compile error with GCC. NFC.
Chandler Carruth [Fri, 19 Aug 2016 09:53:10 +0000 (09:53 +0000)]
[PM] Fix a compile error with GCC. NFC.

llvm-svn: 279228

8 years ago[PM] Make the the new pass manager support fully generic extra arguments
Chandler Carruth [Fri, 19 Aug 2016 09:45:16 +0000 (09:45 +0000)]
[PM] Make the the new pass manager support fully generic extra arguments
to run methods, both for transform passes and analysis passes.

This also allows the analysis manager to use a different set of extra
arguments from the pass manager where useful. Consider passes over
analysis produced units of IR like SCCs of the call graph or loops.
Passes of this nature will often want to refer to the analysis result
that was used to compute their IR units (the call graph or LoopInfo).
And for transformations, they may want to communicate special update
information to the outer pass manager. With this change, it becomes
possible to have a run method for a loop pass that looks more like:

  PreservedAnalyses run(Loop &L, AnalysisManager<Loop, LoopInfo> &AM,
                        LoopInfo &LI, LoopUpdateRecord &UR);

And to query the analysis manager like:

    AM.getResult<MyLoopAnalysis>(L, LI);

This makes accessing the known-available analyses convenient and clear,
and it makes passing customized data structures around easy.

My initial use case is going to be in updating the pass manager layers
when the analysis units of IR change. But there are more use cases here
such as having a layer that lets inner passes signal whether certain
additional passes should be run because of particular simplifications
made. Two desires for this have come up in the past: triggering
additional optimization after successfully unrolling loops, and
triggering additional inlining after collapsing indirect calls to direct
calls.

Despite adding this layer of generic extensibility, the *only* change to
existing, simple usage are for places where we forward declare the
AnalysisManager template. We really shouldn't be doing this because of
the fragility exposed here, but currently it makes coping with the
legacy PM code easier.

Differential Revision: http://reviews.llvm.org/D21462

llvm-svn: 279227

8 years agominor header guards and help messages cleanup
Kirill Bobyrev [Fri, 19 Aug 2016 09:36:14 +0000 (09:36 +0000)]
minor header guards and help messages cleanup

llvm-svn: 279226

8 years ago[PM] Try to work-around what appears to be an MSVC SFINAE issue with
Chandler Carruth [Fri, 19 Aug 2016 09:26:00 +0000 (09:26 +0000)]
[PM] Try to work-around what appears to be an MSVC SFINAE issue with
r279217 where it fails to select the path that other compilers select.

The workaround won't be as careful to produce an error when an analysis
result is incorrect, but we can rely on non-MSVC builds to catch such
errors it seems and MSVC doesn't seem to support the alternative
techniques.

Hoping this brings the windows bots back to life. If not, will have to
revert all of this.

llvm-svn: 279225

8 years ago[CodeGen] Fix a trivial type conversion bug dating back to pre-2008
James Molloy [Fri, 19 Aug 2016 08:38:50 +0000 (08:38 +0000)]
[CodeGen] Fix a trivial type conversion bug dating back to pre-2008

The heuristic above this code is incredibly suspect, but disregarding that it mutates the cast opcode so we need to check the *mutated* opcode later to see if we need to emit an AssertSext or AssertZext node.

Fixes PR29041.

llvm-svn: 279223

8 years ago[asan] Fix size of shadow incorrectly calculated in r279178
Vitaly Buka [Fri, 19 Aug 2016 08:33:53 +0000 (08:33 +0000)]
[asan] Fix size of shadow incorrectly calculated in r279178

Summary: r279178 generates 8 times more stores than necessary.

Reviewers: eugenis

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D23708

llvm-svn: 279222

8 years ago[PM] NFC refactoring: remove the AnalysisManagerBase class, folding it
Chandler Carruth [Fri, 19 Aug 2016 08:31:47 +0000 (08:31 +0000)]
[PM] NFC refactoring: remove the AnalysisManagerBase class, folding it
into the AnalysisManager class template.

Back when I first added this base class there were separate analysis
managers and some plausible reason why it would be a useful factoring of
common code between them. However, after a lot of refactoring cleaning,
we now have *entirely* shared code. The base class was just an arbitrary
division between code in one class template and a separate class
template. It didn't add anything and forced lots of indirection through
"derived_this" for no real gain.

We can always factor a base CRTP class out with common code if there is
ever some *other* analysis manager that wants to share a subset of
logic. But for now, folding things into the primary template is
a non-trivial simplification with no down sides I see. It shortens the
code considerably, removes an unhelpful abstraction, and will make
subsequent patches *dramatically* less complex which enhance the
analysis manager infrastructure to effectively cope with invalidation.

llvm-svn: 279221

8 years ago[ELF] - Do not change binding of symbols when creating relocatable output.
George Rimar [Fri, 19 Aug 2016 08:31:02 +0000 (08:31 +0000)]
[ELF] - Do not change binding of symbols when creating relocatable output.

Spec says "A hidden symbol contained in a relocatable object must be either
removed or converted to STB_LOCAL binding by the link-editor when the
relocatable object is included in an executable file or shared object".
But we previously converted symbols to STB_LOCAL even when -r was specified.

Broken binary was produced, this is PR28967, patch fixes the issue.

Differential revision: https://reviews.llvm.org/D23514

llvm-svn: 279220

8 years ago[modules] Add missing include.
Vassil Vassilev [Fri, 19 Aug 2016 08:30:42 +0000 (08:30 +0000)]
[modules] Add missing include.

llvm-svn: 279219

8 years agoRevert "[CMake] Fix ASM building in llvm/runtimes"
Jonas Hahnfeld [Fri, 19 Aug 2016 08:03:26 +0000 (08:03 +0000)]
Revert "[CMake] Fix ASM building in llvm/runtimes"

This reverts the TSAN parts of commit r279215.

llvm-svn: 279218

8 years ago[PM] Redesign how the new PM detects whether an analysis result provides
Chandler Carruth [Fri, 19 Aug 2016 07:49:23 +0000 (07:49 +0000)]
[PM] Redesign how the new PM detects whether an analysis result provides
its own invalidate method.

Previously, the technique would assume that if a result didn't have an
invalidate method that didn't exactly match the expected signature it
didn't have one at all. This is in fact not the case. And we had
analyses with incorrect signatures for the invalidate method in the
tree that would be erroneously invalidated in certain cases! Yikes.

Moreover a result might legitimately want to have multiple overloads for
the invalidate method, and if one changes or a new one is needed we
again really want a compiler error. For example in the tree we had not
added the overload for a *function* IR unit to the invalidate routine
for TLI. Doh.

So a new techique for the SFINAE detection here: if the result has *any*
member spelled "invalidate" we turn off the synthesis of a default
version. We don't care if it is a member function or a member variable
or how many overloads there are. Once a result has something by that
name it must provide suitable overloads for the contexts in which it is
used. This seems much more resilient and durable.

Huge props to Richard Smith who helped me figure out how on earth we
could even do this in C++. It took quite some doing. The technique is
remarkably clean however, and merely requires that the analysis results
are not *final* classes. I think that's a requirement we can live with
even if it is a bit odd.

I've fixed the two bad in-tree analysis results. And this will make my
next change which changes the API for invalidate much easier to
validate as correct.

llvm-svn: 279217