OSDN Git Service

android-x86/external-llvm.git
7 years ago[sancov] better input parameters validation
Mike Aizatsky [Fri, 3 Mar 2017 18:22:20 +0000 (18:22 +0000)]
[sancov] better input parameters validation

Differential Revision: https://reviews.llvm.org/D30370

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296900 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTry again to appease the FreeBSD bot.
Zachary Turner [Fri, 3 Mar 2017 18:21:04 +0000 (18:21 +0000)]
Try again to appease the FreeBSD bot.

The actual logic was wrong, not just the type conversion.
This should get it correct.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296899 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LoopUnrolling] Peel loops with invariant backedge Phi input
Sanjoy Das [Fri, 3 Mar 2017 18:19:15 +0000 (18:19 +0000)]
[LoopUnrolling] Peel loops with invariant backedge Phi input

Summary:
If a loop contains a Phi node which has an invariant input from back
edge, it is profitable to peel such loops (rather than unroll them) to
use the advantage that this Phi is always invariant starting from 2nd
iteration. After the 1st iteration is peeled, other optimizations can
potentially simplify calculations with this invariant.

Patch by Max Kazantsev!

Reviewers: sanjoy, apilipenko, igor-laevsky, anna, mkuper, reames

Reviewed By: mkuper

Subscribers: mkuper, mzolotukhin, llvm-commits

Differential Revision: https://reviews.llvm.org/D30161

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296898 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LoopUnrolling] Re-prioritize Peeling and Partial unrolling
Sanjoy Das [Fri, 3 Mar 2017 18:19:10 +0000 (18:19 +0000)]
[LoopUnrolling] Re-prioritize Peeling and Partial unrolling

Summary:
In current implementation the loop peeling happens after trip-count based partial unrolling and may
sometimes not happen at all due to it (for example, if trip count is known, but UP.Partial = false). This
is generally bad, the more than there are some situations where peeling is profitable even if the partial
unrolling is disabled.

This patch is a NFC which reorders peeling and partial unrolling application and prepares the code for
implementation of the said optimizations.

Patch by Max Kazantsev!

Reviewers: sanjoy, anna, reames, apilipenko, igor-laevsky, mkuper

Reviewed By: mkuper

Subscribers: mkuper, llvm-commits, mzolotukhin

Differential Revision: https://reviews.llvm.org/D30243

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296897 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] clean up materializeSBB(); NFCI
Sanjay Patel [Fri, 3 Mar 2017 17:58:39 +0000 (17:58 +0000)]
[x86] clean up materializeSBB(); NFCI

This is producing SBB where it is obviously not necessary, so it needs to be limited.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296894 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTry to appease the FreeBSD bots.
Zachary Turner [Fri, 3 Mar 2017 17:56:14 +0000 (17:56 +0000)]
Try to appease the FreeBSD bots.

pthread_self() returns a pthread_t, but we were setting it to
an int.  It seems the cast to int when calling sysctl is still
the correct thing to do, though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296892 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDon't bring in llvm/Support/thread.h in Threading.cpp
Zachary Turner [Fri, 3 Mar 2017 17:39:24 +0000 (17:39 +0000)]
Don't bring in llvm/Support/thread.h in Threading.cpp

Doing so defines the type llvm::thread.  On FreeBSD, we need
to call a macro which references its own ::thread type, which
causes an ambiguity due to ADL when inside of the llvm namespace.

Since we don't even need this unless LLVM_ENABLE_THREADS == 1,
we don't even need this type anyway, as it is always equal to
std::thread, so we can just use that directly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296891 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd #include for unistd.h on Linux.
Zachary Turner [Fri, 3 Mar 2017 17:24:55 +0000 (17:24 +0000)]
Add #include for unistd.h on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296890 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] Provide access to current thread name/thread id.
Zachary Turner [Fri, 3 Mar 2017 17:15:17 +0000 (17:15 +0000)]
[Support] Provide access to current thread name/thread id.

Applications often need the current thread id when making
system calls, and some operating systems provide the notion
of a thread name, which can be useful in enabling better
diagnostics when debugging or logging.

This patch adds an accessor for the thread id, and "best effort"
getters and setters for the thread name.  Since this is
non critical functionality, no error is returned to indicate
that a platform doesn't support thread names.

Differential Revision: https://reviews.llvm.org/D30526

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296887 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse APInt::setBits instead of OR'ing in a separate APInt::getBitsSet call
Simon Pilgrim [Fri, 3 Mar 2017 17:03:52 +0000 (17:03 +0000)]
Use APInt::setBits instead of OR'ing in a separate APInt::getBitsSet call

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296886 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] regenerate checks; NFC
Sanjay Patel [Fri, 3 Mar 2017 16:58:51 +0000 (16:58 +0000)]
[x86] regenerate checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296883 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse APInt::getLowBitsSet instead of APInt::getBitsSet for lower bit mask creation
Simon Pilgrim [Fri, 3 Mar 2017 16:56:33 +0000 (16:56 +0000)]
Use APInt::getLowBitsSet instead of APInt::getBitsSet for lower bit mask creation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296882 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] regenerate checks; NFC
Sanjay Patel [Fri, 3 Mar 2017 16:45:57 +0000 (16:45 +0000)]
[x86] regenerate checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296881 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] regenerate checks; NFC
Sanjay Patel [Fri, 3 Mar 2017 16:42:43 +0000 (16:42 +0000)]
[x86] regenerate checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296880 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse APInt::getOneBitSet instead of APInt::getBitsSet for sign bit mask creation
Simon Pilgrim [Fri, 3 Mar 2017 16:35:57 +0000 (16:35 +0000)]
Use APInt::getOneBitSet instead of APInt::getBitsSet for sign bit mask creation

Avoids all the unnecessary extra bitrange creation/shift stages.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296879 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] regenerate checks; NFC
Sanjay Patel [Fri, 3 Mar 2017 16:34:35 +0000 (16:34 +0000)]
[x86] regenerate checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296877 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] fix formatting; NFC
Sanjay Patel [Fri, 3 Mar 2017 15:17:41 +0000 (15:17 +0000)]
[x86] fix formatting; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296875 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse APInt::getHighBitsSet instead of APInt::getBitsSet for upper bit mask creation
Simon Pilgrim [Fri, 3 Mar 2017 14:37:57 +0000 (14:37 +0000)]
Use APInt::getHighBitsSet instead of APInt::getBitsSet for upper bit mask creation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296874 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU][MC] Fix for Bug 30829 + LIT tests
Dmitry Preobrazhensky [Fri, 3 Mar 2017 14:31:06 +0000 (14:31 +0000)]
[AMDGPU][MC] Fix for Bug 30829 + LIT tests

Added code to check constant bus restrictions for VOP formats (only one SGPR value or literal-constant may be used by the instruction).
Note that the same checks are performed by SIInstrInfo::verifyInstruction (used by lowering code).
Added LIT tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296873 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "Re-apply "[GVNHoist] Move GVNHoist to function simplification part of pipeline.""
Benjamin Kramer [Fri, 3 Mar 2017 14:27:53 +0000 (14:27 +0000)]
Revert "Re-apply "[GVNHoist] Move GVNHoist to function simplification part of pipeline.""

This reverts commit r296759. Miscompiles bash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296872 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse APInt::getOneBitSet instead of APInt::getBitsSet for sign bit mask creation
Simon Pilgrim [Fri, 3 Mar 2017 14:25:46 +0000 (14:25 +0000)]
Use APInt::getOneBitSet instead of APInt::getBitsSet for sign bit mask creation

Avoids all the unnecessary extra bitrange creation/shift stages.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296871 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix Wdocumentation warning
Simon Pilgrim [Fri, 3 Mar 2017 12:09:11 +0000 (12:09 +0000)]
Fix Wdocumentation warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296866 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] fpscr read/write intrinsics not aware of each other
Ranjeet Singh [Fri, 3 Mar 2017 11:40:07 +0000 (11:40 +0000)]
[ARM] fpscr read/write intrinsics not aware of each other

The intrinsics __builtin_arm_get_fpscr and __builtin_arm_set_fpscr read and
write to the fpscr (Floating-Point Status and Control Register) register.

A bug exists in the __builtin_arm_get_fpscr intrinsic definition in llvm which
treats this intrinsic as a IntroNoMem which means it's not a memory access and
doesn't have any other side-effects. Having this property on this intrinsic
means that various optimizations can be done on this such as common
sub-expression elimination with other reads. This can cause issues if there has
been write to this register, e.g.

void foo(int *p) {
     p[0] = __builtin_arm_get_fpscr();
     __builtin_arm_set_fpscr(1);
     p[1] = __builtin_arm_get_fpscr();
}

in the above example the second read is currently CSE'd into the first read,
this is because llvm isn't aware that the write done by __builtin_arm_set_fpscr
effects the same register that __builtin_arm_get_fpscr reads from, to fix this
problem I've removed the property IntrNoMem so that __builtin_arm_get_fpscr is
treated as a memory access.

Differential Revision: https://reviews.llvm.org/D30542

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296865 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Fixes the bug due to absence of in order uses of scalars which needs to be...
Mohammad Shahid [Fri, 3 Mar 2017 10:02:47 +0000 (10:02 +0000)]
[SLP] Fixes the bug due to absence of in order uses of scalars which needs to be available
for VectorizeTree() API.This API uses it for proper mask computation to be used in shufflevector IR.
The fix is to compute the mask for out of order memory accesses while building the vectorizable tree
instead of actual vectorization of vectorizable tree.It also needs to recompute the proper Lane for
external use of vectorizable scalars based on shuffle mask.

Reviewers: mkuper

Differential Revision: https://reviews.llvm.org/D30159

Change-Id: Ide8773ce0ad3562f3cf4d1a0ad0f487e2f60ce5d

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296863 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SDAG] Revert r296476 (and r296486, r296668, r296690).
Chandler Carruth [Fri, 3 Mar 2017 10:02:25 +0000 (10:02 +0000)]
[SDAG] Revert r296476 (and r296486, r296668, r296690).

This patch causes compile times for some patterns to explode. I have
a (large, unreduced) test case that slows down by more than 20x and
several test cases slow down by 2x. I'm sending some of the test cases
directly to Nirav and following up with more details in the review log,
but this should unblock anyone else hitting this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296862 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix a typo in the comments. Patch by marktwtn from https://github.com/llvm-mirror...
Sylvestre Ledru [Fri, 3 Mar 2017 09:36:04 +0000 (09:36 +0000)]
Fix a typo in the comments. Patch by marktwtn from https://github.com/llvm-mirror/llvm/pull/16/files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296860 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Generate VZEROUPPER for Skylake-avx512.
Amjad Aboud [Fri, 3 Mar 2017 09:03:24 +0000 (09:03 +0000)]
[X86] Generate VZEROUPPER for Skylake-avx512.
VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be.

Differential Revision: https://reviews.llvm.org/D29874

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296859 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64AsmParser] rewrite of function parseSysAlias
Sjoerd Meijer [Fri, 3 Mar 2017 08:12:47 +0000 (08:12 +0000)]
[AArch64AsmParser] rewrite of function parseSysAlias

This is a cleanup/rewrite of the parseSysAlias function. It was not using the
tablegen instruction descriptions, but was “manually” matching the mnemonics
and recreating the operands whereas all this information is already in
tablegen; all this code has been replaced with calls to lookupXYZByName
tablegen calls.

Differential Revision: https://reviews.llvm.org/D30491

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296857 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GlobalISel][X86] Support float/double and vector types.
Igor Breger [Fri, 3 Mar 2017 08:06:46 +0000 (08:06 +0000)]
[GlobalISel][X86] Support float/double and vector types.

Summary: [GlobalISel][X86] Add support for f32/f64 and vector types in RegisterBank and InstructionSelector.

Reviewers: delena, zvi

Reviewed By: zvi

Subscribers: dberris, rovka, llvm-commits, kristof.beyls

Differential Revision: https://reviews.llvm.org/D30533

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296856 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r296730, "cmake: Configure the ThinLTO cache directory when using ELF lld...
Peter Collingbourne [Fri, 3 Mar 2017 02:00:22 +0000 (02:00 +0000)]
Revert r296730, "cmake: Configure the ThinLTO cache directory when using ELF lld or gold."

Causes a build failure on the clang-with-thin-lto-ubuntu bot.
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/2117/steps/build-stage3-compiler/logs/stdio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296850 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[msan] Handle x86_sse_stmxcsr and x86_sse_ldmxcsr.
Evgeniy Stepanov [Fri, 3 Mar 2017 01:12:43 +0000 (01:12 +0000)]
[msan] Handle x86_sse_stmxcsr and x86_sse_ldmxcsr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296848 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLiveDebugValues: Assume calls never clobber SP.
Adrian Prantl [Fri, 3 Mar 2017 01:08:25 +0000 (01:08 +0000)]
LiveDebugValues: Assume calls never clobber SP.

A call should never modify the stack pointer, but some backends are
not so sure about this and never list SP in the regmask. For the
purposes of LiveDebugValues we assume a call never clobbers SP. We
already have a similar workaround in DbgValueHistoryCalculator (which
we hopefully can retire soon).

This fixes the availabilty of local ASANified variables on AArch64.

rdar://problem/27757381

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296847 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ProfileData] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Fri, 3 Mar 2017 01:07:34 +0000 (01:07 +0000)]
[ProfileData] Fix some Clang-tidy modernize and Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296846 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCodeGen: BlockPlacement: Precompute layout for chains of triangles.
Kyle Butt [Fri, 3 Mar 2017 01:00:22 +0000 (01:00 +0000)]
CodeGen: BlockPlacement: Precompute layout for chains of triangles.

For chains of triangles with small join blocks that can be tail duplicated, a
simple calculation of probabilities is insufficient. Tail duplication
can be profitable in 3 different ways for these cases:

1) The post-dominators marked 50% are actually taken 56% (This shrinks with
   longer chains)
2) The chains are statically correlated. Branch probabilities have a very
   U-shaped distribution.
   [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
   If the branches in a chain are likely to be from the same side of the
   distribution as their predecessor, but are independent at runtime, this
   transformation is profitable. (Because the cost of being wrong is a small
   fixed cost, unlike the standard triangle layout where the cost of being
   wrong scales with the # of triangles.)
3) The chains are dynamically correlated. If the probability that a previous
   branch was taken positively influences whether the next branch will be
   taken
We believe that 2 and 3 are common enough to justify the small margin in 1.

The code pre-scans a function's CFG to identify this pattern and marks the edges
so that the standard layout algorithm can use the computed results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296845 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[msan] Remove stale comments.
Evgeniy Stepanov [Fri, 3 Mar 2017 00:25:56 +0000 (00:25 +0000)]
[msan] Remove stale comments.

ClStoreCleanOrigin flag was removed back in 2014.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296844 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Fix missing dominator tree dependency
Matt Arsenault [Thu, 2 Mar 2017 23:50:51 +0000 (23:50 +0000)]
AMDGPU: Fix missing dominator tree dependency

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296842 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThinLTOBitcodeWriter: Do not follow operand edges of type GlobalValue when looking...
Peter Collingbourne [Thu, 2 Mar 2017 23:10:17 +0000 (23:10 +0000)]
ThinLTOBitcodeWriter: Do not follow operand edges of type GlobalValue when looking for virtual functions.

Such edges may otherwise result in infinite recursion if a pointer to a vtable
is reachable from the vtable itself. This can happen in practice if a TU
defines the ABI types used to implement RTTI, and is itself compiled with RTTI.

Fixes PR32121.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296839 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMove defClobbersUseOrDef to being a protected member of a class since we don't want...
Daniel Berlin [Thu, 2 Mar 2017 23:06:46 +0000 (23:06 +0000)]
Move defClobbersUseOrDef to being a protected member of a class since we don't want anyone else using it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296838 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[BypassSlowDivision] Use ValueTracking to simplify run-time checks
Nikolai Bozhenov [Thu, 2 Mar 2017 22:12:15 +0000 (22:12 +0000)]
[BypassSlowDivision] Use ValueTracking to simplify run-time checks

ValueTracking is used for more thorough analysis of operands. Based on the
analysis, either run-time checks can be simplified (e.g. check only one operand
instead of two) or the transformation can be avoided. For example, it is quite
often the case that a divisor is promoted from a shorter type and run-time
checks for it are redundant.

With additional compile-time analysis of values, two special cases naturally
arise and are addressed by the patch:

 1) Both operands are known to be short enough. Then, the long division can be
    simply replaced with a short one without CFG modification.

 2) If a division is unsigned and the dividend is known to be short then the
    long division is not needed at all. Because if the divisor is too big for
    short division then the quotient is obviously zero (and the remainder is
    equal to the dividend). Actually, the division is not needed when
    (divisor > dividend).

Differential Revision: https://reviews.llvm.org/D29897

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296832 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCMake: Clean up VersionFromVCS.cmake
Tom Stellard [Thu, 2 Mar 2017 22:05:13 +0000 (22:05 +0000)]
CMake: Clean up VersionFromVCS.cmake

Summary:
Fix a few problems in VersionFromVCS.cmake to make it more reliable:

- Stop using git svn info to retrieve the svn revision.  I am unable to
  determine what the svn revision returned by this command means.
  During my testing this command returned a revision from a month
  ago which was not the HEAD of any of my local branches.

  Also, this revision was never actually added to the version string due
  to a typo in the script.  All it was used for was to reject the
  revision number returned by git svn find-rev HEAD when the revision
  numbers didn't match.

- Populate GIT_COMMIT even when we detect a git repo without any
  svn information.

Reviewers: mehdi_amini, beanz

Reviewed By: beanz

Subscribers: mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D30092

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296829 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[BypassSlowDivision] Refactor fast division insertion logic (NFC)
Nikolai Bozhenov [Thu, 2 Mar 2017 22:05:07 +0000 (22:05 +0000)]
[BypassSlowDivision] Refactor fast division insertion logic (NFC)

The most important goal of the patch is to break large insertFastDiv function
into separate pieces, so that later a different fast insertion logic can be
implemented using some of these pieces.

Differential Revision: https://reviews.llvm.org/D29896

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296828 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] Fix DebugLoc propagation when folding !(x cc y) -> (x !cc y)
Taewook Oh [Thu, 2 Mar 2017 21:58:35 +0000 (21:58 +0000)]
[DAGCombiner] Fix DebugLoc propagation when folding !(x cc y) -> (x !cc y)

Summary:
Currently, when 't1: i1 = setcc t2, t3, cc' followed by 't4: i1 = xor t1, Constant:i1<-1>' is folded into 't5: i1 = setcc t2, t3 !cc', SDLoc of newly created SDValue 't5' follows SDLoc of 't4', not 't1'. However, as the opcode of newly created SDValue is 'setcc', it make more sense to take DebugLoc from 't1' than 't4'. For the code below

```
extern int bar();
extern int baz();

int foo(int x, int y) {
  if (x != y)
    return bar();
  else
    return baz();
}
```

, following is the bitcode representation of 'foo' at the end of llvm-ir level optimization:

```
define i32 @foo(i32 %x, i32 %y) !dbg !4 {
entry:
  tail call void @llvm.dbg.value(metadata i32 %x, i64 0, metadata !9, metadata !11), !dbg !12
  tail call void @llvm.dbg.value(metadata i32 %y, i64 0, metadata !10, metadata !11), !dbg !13
  %cmp = icmp ne i32 %x, %y, !dbg !14
  br i1 %cmp, label %if.then, label %if.else, !dbg !16

if.then:                                          ; preds = %entry
  %call = tail call i32 (...) @bar() #3, !dbg !17
  br label %return, !dbg !18

if.else:                                          ; preds = %entry
  %call1 = tail call i32 (...) @baz() #3, !dbg !19
  br label %return, !dbg !20

return:                                           ; preds = %if.else, %if.then
  %retval.0 = phi i32 [ %call, %if.then ], [ %call1, %if.else ]
  ret i32 %retval.0, !dbg !21
}

!14 = !DILocation(line: 5, column: 9, scope: !15)
!16 = !DILocation(line: 5, column: 7, scope: !4)

```

As you can see, in 'entry' block, 'icmp' instruction and 'br' instruction have different debug locations. However, with current implementation, there's no distinction between debug locations of these two when they are lowered to asm instructions. This is because 'icmp' and 'br' become 'setcc' 'xor' and 'brcond' in SelectionDAG, where SDLoc of 'setcc' follows the debug location of 'icmp' but SDLOC of 'xor' and 'brcond' follows the debug location of 'br' instruction, and SDLoc of 'xor' overwrites SDLoc of 'setcc' when they are folded. This patch addresses this issue.

Reviewers: atrick, bogner, andreadb, craig.topper, aprantl

Reviewed By: andreadb

Subscribers: jlebar, mkuper, jholewinski, andreadb, llvm-commits

Differential Revision: https://reviews.llvm.org/D29813

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296825 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAG] early exit to improve readability and formatting of visitMemCmpCall(); NFCI
Sanjay Patel [Thu, 2 Mar 2017 21:56:43 +0000 (21:56 +0000)]
[DAG] early exit to improve readability and formatting of visitMemCmpCall(); NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296824 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Pick the right branch opcode depending on branch probabilities
Krzysztof Parzyszek [Thu, 2 Mar 2017 21:49:49 +0000 (21:49 +0000)]
[Hexagon] Pick the right branch opcode depending on branch probabilities

Specifically, pick the opcode with the correct branch prediction, i.e.
jump:t or jump:nt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296821 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "AMDGPU: Re-do update for branch-relaxation test"
Tobias Grosser [Thu, 2 Mar 2017 21:47:51 +0000 (21:47 +0000)]
Revert "AMDGPU: Re-do update for branch-relaxation test"

This commit also relied on r296812, which I just reverted. We should probably
apply it again, after the r296812 has been discussed and been reapplied in some
variant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296820 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCodeGen: MachineBlockPlacement: Remove the unused outlining heuristic.
Kyle Butt [Thu, 2 Mar 2017 21:44:24 +0000 (21:44 +0000)]
CodeGen: MachineBlockPlacement: Remove the unused outlining heuristic.

Outlining optional branches isn't a good heuristic, and it's never been
on by default. Remove it to clean things up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296818 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Fix insert point for store rescheduling.
Eli Friedman [Thu, 2 Mar 2017 21:39:39 +0000 (21:39 +0000)]
[ARM] Fix insert point for store rescheduling.

In ARMPreAllocLoadStoreOpt::RescheduleOps, LastOp should be the last
operation which we want to merge. If we break out of the loop because
an operation has the wrong offset, we shouldn't use that operation
as LastOp.

This patch fixes some cases where we would move stores to the wrong
insert point.

Re-commit with a fix to increment NumMove in the right place.

Differential Revision: https://reviews.llvm.org/D30124

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296815 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "Fix PR 24415 (at least), by making our post-dominator tree behavior sane."
Tobias Grosser [Thu, 2 Mar 2017 21:08:37 +0000 (21:08 +0000)]
Revert "Fix PR 24415 (at least), by making our post-dominator tree behavior sane."

and also "clang-format GenericDomTreeConstruction.h, since the current
formatting makes it look like their is a bug in the loop indentation, and there
is not"

This reverts commit r296535.

There are still some open design questions which I would like to discuss. I
revert this for Daniel (who gave the OK), as he is on vacation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296812 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PPC] Fix code generation for bswap(int32) followed by store16
Guozhi Wei [Thu, 2 Mar 2017 21:07:59 +0000 (21:07 +0000)]
[PPC] Fix code generation for bswap(int32) followed by store16

This patch fixes pr32063.

Current code in PPCTargetLowering::PerformDAGCombine can transform

bswap
store

into a single PPCISD::STBRX instruction. but it doesn't consider the case that the operand size of bswap may be larger than store size. When it occurs, we need 2 modifications,

1 For the last operand of PPCISD::STBRX, we should not use DAG.getValueType(N->getOperand(1).getValueType()), instead we should use cast<StoreSDNode>(N)->getMemoryVT().

2 Before PPCISD::STBRX, we need to shift the original operand of bswap to the right side.

Differential Revision: https://reviews.llvm.org/D30362

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296811 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] Move Stream library from MSF -> Support.
Zachary Turner [Thu, 2 Mar 2017 20:52:51 +0000 (20:52 +0000)]
[Support] Move Stream library from MSF -> Support.

After several smaller patches to get most of the core improvements
finished up, this patch is a straight move and header fixup of
the source.

Differential Revision: https://reviews.llvm.org/D30266

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296810 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Extend redundant copy elimination pass to handle non-zero stores.
Chad Rosier [Thu, 2 Mar 2017 20:48:11 +0000 (20:48 +0000)]
[AArch64] Extend redundant copy elimination pass to handle non-zero stores.

This patch extends the current functionality of the AArch64 redundant copy
elimination pass to handle non-zero cases such as:

BB#0:
  cmp x0, #1
  b.eq .LBB0_1
.LBB0_1:
  orr x0, xzr, #0x1  ; <-- redundant copy; x0 known to hold #1.

Differential Revision: https://reviews.llvm.org/D29344

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296809 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAG] improve documentation comments; NFC
Sanjay Patel [Thu, 2 Mar 2017 20:48:08 +0000 (20:48 +0000)]
[DAG] improve documentation comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296808 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MSP430] Add SRet support to MSP430 target
Vadzim Dambrouski [Thu, 2 Mar 2017 20:25:10 +0000 (20:25 +0000)]
[MSP430] Add SRet support to MSP430 target

This patch adds support for struct return values to the MSP430
target backend. It also reverses the order of argument and return
registers in the calling convention to bring it into closer
alignment with the published EABI from TI.

Patch by Andrew Wygle (awygle).

Differential Revision: https://reviews.llvm.org/D29069

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296807 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThe patch fixes r296770
Evgeny Stupachenko [Thu, 2 Mar 2017 19:41:38 +0000 (19:41 +0000)]
The patch fixes r296770
Summary:

Extend -unroll-partial-threshold to 200 for runtime-loop3.ll test
as epilogue unroll initially add 1 more IV to the loop.

From: Evgeny Stupachenko <evstupac@gmail.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296803 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Reduce amount of boilerplate code used to select load instruction opcode.
Artem Belevich [Thu, 2 Mar 2017 19:14:14 +0000 (19:14 +0000)]
[NVPTX] Reduce amount of boilerplate code used to select load instruction opcode.

Make opcode selection code for the load instruction a bit easier
to read and maintain.

This patch also catches number of f16 load/store variants that were
not handled before.

Differential Revision: https://reviews.llvm.org/D30513

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296785 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Added missing LDU/LDG intrinsics for f16.
Artem Belevich [Thu, 2 Mar 2017 19:14:10 +0000 (19:14 +0000)]
[NVPTX] Added missing LDU/LDG intrinsics for f16.

Differential Revision: https://reviews.llvm.org/D30512

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296784 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix some Wdocumentation warnings
Simon Pilgrim [Thu, 2 Mar 2017 18:59:07 +0000 (18:59 +0000)]
Fix some Wdocumentation warnings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296783 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][MMX] Fixed i32 extraction on 32-bit targets
Simon Pilgrim [Thu, 2 Mar 2017 18:56:06 +0000 (18:56 +0000)]
[X86][MMX] Fixed i32 extraction on 32-bit targets

MMX extraction often ends up as extract_i32(bitcast_v2i32(extract_i64(bitcast_v1i64(x86mmx v), 0)), 0) which fails to simplify on 32-bit targets as i64 isn't legal

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296782 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCast to the right type on Windows.
Vassil Vassilev [Thu, 2 Mar 2017 18:12:59 +0000 (18:12 +0000)]
Cast to the right type on Windows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296778 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Skip blocks that define vector predicate registers in early-if
Krzysztof Parzyszek [Thu, 2 Mar 2017 18:10:59 +0000 (18:10 +0000)]
[Hexagon] Skip blocks that define vector predicate registers in early-if

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296777 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove redundant include.
Vassil Vassilev [Thu, 2 Mar 2017 18:04:44 +0000 (18:04 +0000)]
Remove redundant include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296775 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoReland r296442 with modifications reverted in r296463.
Vassil Vassilev [Thu, 2 Mar 2017 17:56:45 +0000 (17:56 +0000)]
Reland r296442 with modifications reverted in r296463.

Original commit message:

"Allow externally dlopen-ed libraries to be registered as permanent libraries.

This is also useful in cases when llvm is in a shared library. First we dlopen
the llvm shared library and then we register it as a permanent library in order
to keep the JIT and other services working.

Patch reviewed by Vedant Kumar (D29955)!"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296774 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Properly handle 'q' constraint in 128-byte vector mode
Krzysztof Parzyszek [Thu, 2 Mar 2017 17:50:24 +0000 (17:50 +0000)]
[Hexagon] Properly handle 'q' constraint in 128-byte vector mode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296772 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PowerPC][ELFv2ABI] Allocate parameter area on-demand to reduce stack frame size
Nemanja Ivanovic [Thu, 2 Mar 2017 17:38:59 +0000 (17:38 +0000)]
[PowerPC][ELFv2ABI] Allocate parameter area on-demand to reduce stack frame size

This patch reduces the stack frame size by not allocating the parameter area if
it is not required. In the current implementation LowerFormalArguments_64SVR4
already handles the parameter area, but LowerCall_64SVR4 does not
(when calculating the stack frame size). What this patch does is make
LowerCall_64SVR4 consistent with LowerFormalArguments_64SVR4.

Committing on behalf of Hiroshi Inoue.

Differential Revision: https://reviews.llvm.org/D29881

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296771 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThe patch turns on epilogue unroll for loops with constant recurency start.
Evgeny Stupachenko [Thu, 2 Mar 2017 17:38:46 +0000 (17:38 +0000)]
The patch turns on epilogue unroll for loops with constant recurency start.
Summary:

Set unroll remainder to epilog if a loop contains a phi with constant parameter:

  loop:
  pn = phi [Const, PreHeader], [pn.next, Latch]
  ...

Reviewer: hfinkel

Differential Revision: http://reviews.llvm.org/D27004

From: Evgeny Stupachenko <evstupac@gmail.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296770 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] avoid assertion when folding binops with opaque constants
Sanjay Patel [Thu, 2 Mar 2017 17:18:56 +0000 (17:18 +0000)]
[DAGCombiner] avoid assertion when folding binops with opaque constants

This bug was introduced with:
https://reviews.llvm.org/rL296699

There may be a way to loosen the restriction, but for now just bail out
on any opaque constant.

The tests show that opacity is target-specific. This goes back to cost
calculations in ConstantHoisting based on TTI->getIntImmCost().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296768 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoNew tool: opt-diff.py
Adam Nemet [Thu, 2 Mar 2017 17:00:59 +0000 (17:00 +0000)]
New tool: opt-diff.py

This tool allows generating the different between two optimization record
files.  The result is a YAML file too that can be visualized with opt-viewer.

This is very useful to see what optimization were added and removed by a
change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296767 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[opt-viewer] Treat remarks with different attributes as different
Adam Nemet [Thu, 2 Mar 2017 17:00:56 +0000 (17:00 +0000)]
[opt-viewer] Treat remarks with different attributes as different

We used to exclude arguments but for a diffed YAML file, it's interesting to
show these as changes.

Turns out this also affects gvn/LoadClobbered because we used to squash
multiple entries of this on the same line even if they reported clobbers
by *different* instructions.  This increases the number of unique entries now
and the share of gvn/LoadClobbered.

Total number of remarks      902287

Top 10 remarks by pass:
  inline                         43%
  gvn                            37%
  licm                           11%
  loop-vectorize                  4%
  asm-printer                     3%
  regalloc                        1%
  loop-unroll                     1%
  inline-cost                     0%
  slp-vectorizer                  0%
  loop-delete                     0%

Top 10 remarks:
  gvn/LoadClobbered              33%
  inline/Inlined                 16%
  inline/CanBeInlined            14%
  inline/NoDefinition             7%
  licm/Hoisted                    6%
  licm/LoadWithLoopInvariantAddressInvalidated  5%
  gvn/LoadElim                    3%
  asm-printer/InstructionCount    3%
  inline/TooCostly                2%
  loop-vectorize/MissedDetails    2%

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296766 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[opt-viewer] Don't use __getattr__ for missing YAML attributes
Adam Nemet [Thu, 2 Mar 2017 17:00:53 +0000 (17:00 +0000)]
[opt-viewer] Don't use __getattr__ for missing YAML attributes

__getattr__ does not work well with debugging.  If the attribute function has
a run-time error, a missing attribute is reported instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296765 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[opt-viewer] Sort entries with identical hotness by source line
Adam Nemet [Thu, 2 Mar 2017 17:00:49 +0000 (17:00 +0000)]
[opt-viewer] Sort entries with identical hotness by source line

We want entries that are close to each other in the source appear next to each
other.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296764 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAllow use of spaces in Bugpoint ‘--compile-command’ argument
David Bozier [Thu, 2 Mar 2017 16:50:48 +0000 (16:50 +0000)]
Allow use of spaces in Bugpoint ‘--compile-command’ argument

Bug-Point functionality needs extending due to the patch D29185 by bd1976llvm (Allow llvm's build and test systems to support paths with spaces ). It requires Bugpoint to accept the use of spaces within ‘--compile-command’ tokens.

Details
Bugpoint uses the argument ‘--compile-command’ to pass in a command line argument as a string, the string is tokenized by the ‘lexCommand’ function using spaces as a delimiter. Patch D29185 will cause the unit test compile-custom.ll to fail as spaces are now required within tokens and as a delimiter. This patch allows the use of escape characters as below:

Two consecutive '\' evaluate to a single '\'.
A space after a '\' evaluates to a space that is not interpreted as a delimiter.
Any other instances of the '\' character are removed.

Committed on behalf of Owen Reynolds

Differential revision: https://reviews.llvm.org/D29940

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296763 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agofix typo in comment; NFC
Sanjay Patel [Thu, 2 Mar 2017 16:37:24 +0000 (16:37 +0000)]
fix typo in comment; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296760 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRe-apply "[GVNHoist] Move GVNHoist to function simplification part of pipeline."
Geoff Berry [Thu, 2 Mar 2017 16:16:47 +0000 (16:16 +0000)]
Re-apply "[GVNHoist] Move GVNHoist to function simplification part of pipeline."

This re-applies r289696, which caused TSan perf regression, which has
since been addressed in separate changes (see PR for details).

See PR31382.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296759 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGlobalISel: record correct stack usage for signext parameters.
Tim Northover [Thu, 2 Mar 2017 15:34:18 +0000 (15:34 +0000)]
GlobalISel: record correct stack usage for signext parameters.

The CallingConv.td rules allocate 8 bytes for these kinds of arguments
on AAPCS targets, but we were only recording the smaller amount. The
difference is theoretical on AArch64 because we don't actually store
more than the smaller amount, but it's still much better to have these
two components in agreement.

Based on Diana Picus's ARM equivalent patch (where it matters a lot
more).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296754 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] Avoid faulty combines of select-cmp-br
Bjorn Pettersson [Thu, 2 Mar 2017 15:18:58 +0000 (15:18 +0000)]
[InstCombine] Avoid faulty combines of select-cmp-br

Summary:
When InstCombine is optimizing certain select-cmp-br patterns
it replaces the result of the select in uses outside of the
basic block containing the select. This is only legal if the
path from the select to the outside use is disjoint from all
other paths out from the originating basic block.

The problem found was that InstCombiner::replacedSelectWithOperand
did not consider the case when both edges out from the br pointed
to the same label. In that case the paths aren't disjoint and the
transformation is illegal. This patch avoids the faulty rewrites
by verifying that there is a single flow to the successor where
we want to replace uses.

Reviewers: llvm-commits, spatel, majnemer

Differential Revision: https://reviews.llvm.org/D30455

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296752 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM/AArch64] Update costs for interleaved accesses with wide types
Matthew Simpson [Thu, 2 Mar 2017 15:15:35 +0000 (15:15 +0000)]
[ARM/AArch64] Update costs for interleaved accesses with wide types

After r296750, we're able to match interleaved accesses having types wider than
128 bits. This patch updates the associated TTI costs.

Differential Revision: https://reviews.llvm.org/D29675

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296751 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM/AArch64] Support wide interleaved accesses
Matthew Simpson [Thu, 2 Mar 2017 15:11:20 +0000 (15:11 +0000)]
[ARM/AArch64] Support wide interleaved accesses

This patch teaches (ARM|AArch64)ISelLowering.cpp to match illegal vector types
to interleaved access intrinsics as long as the types are multiples of the
vector register width. A "wide" access will now be mapped to multiple
interleave intrinsics similar to the way in which non-interleaved accesses with
illegal types are legalized into multiple accesses. I'll update the associated
TTI costs (in getInterleavedMemoryOpCost) as a follow-on.

Differential Revision: https://reviews.llvm.org/D29466

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296750 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDo not leak OpenedHandles.
Vassil Vassilev [Thu, 2 Mar 2017 14:30:05 +0000 (14:30 +0000)]
Do not leak OpenedHandles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296748 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LV] Considier non-consecutive but vectorizable accesses for VF selection
Matthew Simpson [Thu, 2 Mar 2017 13:55:05 +0000 (13:55 +0000)]
[LV] Considier non-consecutive but vectorizable accesses for VF selection

When computing the smallest and largest types for selecting the maximum
vectorization factor, we currently ignore loads and stores of pointer types if
the memory access is non-consecutive. We do this because such accesses must be
scalarized regardless of vectorization factor, and thus shouldn't be considered
when determining the factor. This patch makes this check less aggressive by
also considering non-consecutive accesses that may be vectorized, such as
interleaved accesses. Because we don't know at the time of the check if an
accesses will certainly be vectorized (this is a cost model decision given a
particular VF), we consider all accesses that can potentially be vectorized.

Differential Revision: https://reviews.llvm.org/D30305

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296747 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdded special test covering a problem with PIC relocation model on SLM architecture...
Andrew V. Tischenko [Thu, 2 Mar 2017 13:47:03 +0000 (13:47 +0000)]
Added special test covering a problem with PIC relocation model on SLM architecture. The fix will come in D26855.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296746 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDo not verify MachimeDominatorTree if it is not calculated
Serge Pavlov [Thu, 2 Mar 2017 12:00:10 +0000 (12:00 +0000)]
Do not verify MachimeDominatorTree if it is not calculated

If dominator tree is not calculated or is invalidated, set corresponding
pointer in the pass state to nullptr. Such pointer value will indicate
that operations with dominator tree are not allowed. In particular, it
allows to skip verification for such pass state. The dominator tree is
not calculated if the machine dominator pass was skipped, it occures in
the case of entities with linkage available_externally.

The change fixes some test fails observed when expensive checks
are enabled.

Differential Revision: https://reviews.llvm.org/D29280

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296742 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix typo. NFCI
Xin Tong [Thu, 2 Mar 2017 08:39:11 +0000 (08:39 +0000)]
Fix typo. NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296735 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agocmake: Configure the ThinLTO cache directory when using ELF lld or gold.
Peter Collingbourne [Thu, 2 Mar 2017 03:01:12 +0000 (03:01 +0000)]
cmake: Configure the ThinLTO cache directory when using ELF lld or gold.

Differential Revision: https://reviews.llvm.org/D30522

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296730 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLTO: When creating a local cache, create the cache directory if it does not already...
Peter Collingbourne [Thu, 2 Mar 2017 02:02:38 +0000 (02:02 +0000)]
LTO: When creating a local cache, create the cache directory if it does not already exist.

Differential Revision: https://reviews.llvm.org/D30519

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296726 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLiveRegMatrix: Fix some subreg interference checks
Matthias Braun [Thu, 2 Mar 2017 00:35:08 +0000 (00:35 +0000)]
LiveRegMatrix: Fix some subreg interference checks

Surprisingly, one of the three interference checks in LiveRegMatrix was
using the main live range instead of the apropriate subregister range
resulting in unnecessarily conservative results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296722 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLiveIntervalUnion: Remove unused function; NFC
Matthias Braun [Thu, 2 Mar 2017 00:15:06 +0000 (00:15 +0000)]
LiveIntervalUnion: Remove unused function; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296721 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r296708; causing test failures on ARM hosts.
Eli Friedman [Thu, 2 Mar 2017 00:08:50 +0000 (00:08 +0000)]
Revert r296708; causing test failures on ARM hosts.

Original commit message:

[ARM] Fix insert point for store rescheduling.

In ARMPreAllocLoadStoreOpt::RescheduleOps, LastOp should be the last
operation which we want to merge. If we break out of the loop because
an operation has the wrong offset, we shouldn't use that operation as
LastOp.

This patch fixes some cases where we would sink stores for no reason.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296718 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Wed, 1 Mar 2017 23:59:26 +0000 (23:59 +0000)]
[Support] Fix some Clang-tidy modernize and Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296714 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove spurious use of LLVM_FALLTHROUGH (NFC)
Paul Robinson [Wed, 1 Mar 2017 23:59:11 +0000 (23:59 +0000)]
Remove spurious use of LLVM_FALLTHROUGH (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296713 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] mulhi + 1 never overflow.
Amaury Sechet [Wed, 1 Mar 2017 23:44:17 +0000 (23:44 +0000)]
[DAGCombiner] mulhi + 1 never overflow.

Summary:
This can be used to optimize large multiplications after legalization.

Depends on D29565

Reviewers: mkuper, spatel, RKSimon, zvi, bkramer, aaboud, craig.topper

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D29587

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296711 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GlobalISel] Add a way for targets to enable GISel.
Ahmed Bougacha [Wed, 1 Mar 2017 23:33:08 +0000 (23:33 +0000)]
[GlobalISel] Add a way for targets to enable GISel.

Until now, we've had to use -global-isel to enable GISel.  But using
that on other targets that don't support it will result in an abort, as we
can't build a full pipeline.
Additionally, we want to experiment with enabling GISel by default for
some targets: we can't just enable GISel by default, even among those
target that do have some support, because the level of support varies.

This first step adds an override for the target to explicitly define its
level of support.  For AArch64, do that using
a new command-line option (I know..):
  -aarch64-enable-global-isel-at-O=<N>
Where N is the opt-level below which GISel should be used.

Default that to -1, so that we still don't enable GISel anywhere.
We're not there yet!

While there, remove a couple LLVM_UNLIKELYs.  Building the pipeline is
such a cold path that in practice that shouldn't matter at all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296710 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoImprove mulhi overflow test. NFC
Amaury Sechet [Wed, 1 Mar 2017 23:31:19 +0000 (23:31 +0000)]
Improve mulhi overflow test. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296709 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Fix insert point for store rescheduling.
Eli Friedman [Wed, 1 Mar 2017 23:20:29 +0000 (23:20 +0000)]
[ARM] Fix insert point for store rescheduling.

In ARMPreAllocLoadStoreOpt::RescheduleOps, LastOp should be the last
operation which we want to merge. If we break out of the loop because
an operation has the wrong offset, we shouldn't use that operation as
LastOp.

This patch fixes some cases where we would sink stores for no reason.

Differential Revision: https://reviews.llvm.org/D30124

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296708 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd polly to svn:ignore.
Eli Friedman [Wed, 1 Mar 2017 23:16:35 +0000 (23:16 +0000)]
Add polly to svn:ignore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296705 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Check correct instructions for load/store rescheduling.
Eli Friedman [Wed, 1 Mar 2017 22:56:20 +0000 (22:56 +0000)]
[ARM] Check correct instructions for load/store rescheduling.

This code starts from the high end of the sorted vector of offsets, and
works backwards: it tries to find contiguous offsets, process them, then
pops them from the end of the vector. Most of the code agrees with this
order of processing, but one loop doesn't: it instead processes elements
from the low end of the vector (which are nodes with unrelated offsets).
Fix that loop to process the correct elements.

This has a few implications. One, we don't incorrectly return early when
processing multiple groups of offsets in the same block (which allows
rescheduling prera-ldst-insertpt.mir). Two, we pick the correct insert
point for loads, so they're correctly sorted (which affects the
scheduling of vldm-liveness.ll). I think it might also impact some of
the heuristics slightly.

Differential Revision: https://reviews.llvm.org/D30368

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296701 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] fold binops with constant into select-of-constants
Sanjay Patel [Wed, 1 Mar 2017 22:51:31 +0000 (22:51 +0000)]
[DAGCombiner] fold binops with constant into select-of-constants

This is part of the ongoing attempt to improve select codegen for all targets and select
canonicalization in IR (see D24480 for more background). The transform is a subset of what
is done in InstCombine's FoldOpIntoSelect().

I first noticed a regression in the x86 avx512-insert-extract.ll tests with a patch that
hopes to convert more selects to basic math ops. This appears to be a general missing DAG
transform though, so I added tests for all standard binops in rL296621
(PowerPC was chosen semi-randomly; it has scripted FileCheck support, but so do ARM and x86).

The poor output for "sel_constants_shl_constant" is tracked with:
https://bugs.llvm.org/show_bug.cgi?id=32105

Differential Revision: https://reviews.llvm.org/D30502

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296699 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Constant Hoisting] Avoid inserting instructions before EH pads
Reid Kleckner [Wed, 1 Mar 2017 22:41:12 +0000 (22:41 +0000)]
[Constant Hoisting] Avoid inserting instructions before EH pads

Now that terminators can be EH pads, this code needs to iterate over the
immediate dominators of the EH pad to find a valid insertion point.

Fix for PR32107

Patch by Robert Olliff!

Differential Revision: https://reviews.llvm.org/D30511

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296698 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MC] Fix MachineLocation constructor broken in r294685 (NFC).
Eugene Zelenko [Wed, 1 Mar 2017 22:28:23 +0000 (22:28 +0000)]
[MC] Fix MachineLocation constructor broken in r294685 (NFC).

Problem spotted by Frej Drejhammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296697 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd test case for mulhi's overflow. NFC
Amaury Sechet [Wed, 1 Mar 2017 22:27:21 +0000 (22:27 +0000)]
Add test case for mulhi's overflow. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296696 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DebugInfo] [DWARFv5] Unique abbrevs for DIEs with different implicit_const values
Victor Leschuk [Wed, 1 Mar 2017 22:13:42 +0000 (22:13 +0000)]
[DebugInfo] [DWARFv5] Unique abbrevs for DIEs with different implicit_const values

Take DW_FORM_implicit_const attribute value into account when profiling
DIEAbbrevData.

Currently if we have two similar types with implicit_const attributes and
different values we end up with only one abbrev in .debug_abbrev section.
For example consider two structures: S1 with implicit_const attribute ATTR
and value VAL1 and S2 with implicit_const ATTR and value VAL2.
The .debug_abbrev section will contain only 1 related record:

[N] DW_TAG_structure_type       DW_CHILDREN_yes
        DW_AT_ATTR        DW_FORM_implicit_const  VAL1
        // ....

This is incorrect as struct S2 (with VAL2) will use abbrev record with VAL1.

With this patch we will have two different abbreviations here:

[N] DW_TAG_structure_type       DW_CHILDREN_yes
        DW_AT_ATTR        DW_FORM_implicit_const  VAL1
        // ....

[M] DW_TAG_structure_type       DW_CHILDREN_yes
        DW_AT_ATTR        DW_FORM_implicit_const  VAL2
        // ....

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296691 91177308-0d34-0410-b5e6-96231b3b80d8