OSDN Git Service

android-x86/external-llvm.git
8 years agoAMDGPU/SI: Correctly initialize SIInsertWaits pass
Tom Stellard [Fri, 5 Feb 2016 17:42:38 +0000 (17:42 +0000)]
AMDGPU/SI: Correctly initialize SIInsertWaits pass

Reviewers: arsenm

Subscribers: arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D16724

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259894 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[WebAssembly] Update the select instructions' operand orders to match the spec.
Dan Gohman [Fri, 5 Feb 2016 17:14:59 +0000 (17:14 +0000)]
[WebAssembly] Update the select instructions' operand orders to match the spec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259893 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd the missing test case for PR26193
Nemanja Ivanovic [Fri, 5 Feb 2016 15:03:17 +0000 (15:03 +0000)]
Add the missing test case for PR26193

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259888 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix for PR 26193
Nemanja Ivanovic [Fri, 5 Feb 2016 14:50:29 +0000 (14:50 +0000)]
Fix for PR 26193

This is a simple fix for a PowerPC intrinsic that was incorrectly defined
(the return type was incorrect).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259886 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMove classes defined in a cpp file into an anonymous namespace.
Benjamin Kramer [Fri, 5 Feb 2016 13:50:53 +0000 (13:50 +0000)]
Move classes defined in a cpp file into an anonymous namespace.

No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259883 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoPrefix external symbols in llvm-c-test.
Benjamin Kramer [Fri, 5 Feb 2016 13:31:14 +0000 (13:31 +0000)]
Prefix external symbols in llvm-c-test.

This makes it less likely to clash with other stuff that might be linked
in by change, e.g. ncurses exposes an external function called simply
"echo", so linking ncurses statically into the binary explodes in funny
ways.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259882 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "[AArch64] Improve load/store optimizer to handle LDUR + LDR (take 3)."
Renato Golin [Fri, 5 Feb 2016 12:14:30 +0000 (12:14 +0000)]
Revert "[AArch64] Improve load/store optimizer to handle LDUR + LDR (take 3)."

This reverts commit r259812 as it broke AArch64 self-hosting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259881 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[DebugInfo] Eliminate compilation warning about used variable LSDA
Dmitry Polukhin [Fri, 5 Feb 2016 09:24:34 +0000 (09:24 +0000)]
[DebugInfo] Eliminate compilation warning about used variable LSDA

The waring was:
lib/DebugInfo/DWARF/DWARFDebugFrame.cpp:643:20: warning: variable ‘LSDA’ set but not used

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259877 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LoopUnrolling] Try harder to avoid rebuilding LCSSA when possible.
Michael Zolotukhin [Fri, 5 Feb 2016 02:17:36 +0000 (02:17 +0000)]
[LoopUnrolling] Try harder to avoid rebuilding LCSSA when possible.

In r255133 (reapplied r253126) we started to avoid redundant
recomputation of LCSSA after loop-unrolling. This patch moves one step
further in this direction - now we can avoid it for much wider range of
loops, as we start to look at IR and try to figure out if the
transformation actually breaks LCSSA phis or makes it necessary to
insert new ones.

Differential Revision: http://reviews.llvm.org/D16838

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259869 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[MC] Add support for encoding CodeView variable definition ranges
David Majnemer [Fri, 5 Feb 2016 01:55:49 +0000 (01:55 +0000)]
[MC] Add support for encoding CodeView variable definition ranges

CodeView, like most other debug formats, represents the live range of a
variable so that debuggers might print them out.

They use a variety of records to represent how a particular variable
might be available (in a register, in a frame pointer, etc.) along with
a set of ranges where this debug information is relevant.

However, the format only allows us to use ranges which are limited to a
maximum of 0xF000 in size.  This means that we need to split our debug
information into chunks of 0xF000.

Because the layout of code is not known until *very* late, we must use a
new fragment to record the information we need until we can know
*exactly* what the range is.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259868 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[RS4GC] Pass DenseMap by reference, NFC
Joseph Tremoulet [Fri, 5 Feb 2016 01:42:52 +0000 (01:42 +0000)]
[RS4GC] Pass DenseMap by reference, NFC

Summary:
Passing the rematerialized values map to insertRematerializationStores by
value looks to be a simple oversight; update it to pass by reference.

Reviewers: reames, sanjoy

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D16911

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259867 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd various binary operations in the LLVM C API echo test
Amaury Sechet [Fri, 5 Feb 2016 01:27:11 +0000 (01:27 +0000)]
Add various binary operations in the LLVM C API echo test

Summary: This diff increase the tested surface of the C API.

Reviewers: bogner, chandlerc, echristo, dblaikie, joker.eph, Wallbraker

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D16910

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259863 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LoopLoadElim] Don't allow versioning when optForSize
Adam Nemet [Fri, 5 Feb 2016 01:14:05 +0000 (01:14 +0000)]
[LoopLoadElim] Don't allow versioning when optForSize

This was requested in the review of D16300.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259861 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix typo in comment
Adam Nemet [Fri, 5 Feb 2016 01:14:00 +0000 (01:14 +0000)]
Fix typo in comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259860 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix printing of f16 machine operands
Matt Arsenault [Fri, 5 Feb 2016 00:50:18 +0000 (00:50 +0000)]
Fix printing of f16 machine operands

Only single and double FP immediates are correctly printed by
MachineInstr::print() during debug output. Half float type goes to
APFloat::convertToDouble() and hits assertion it is not a double
semantics. This diff prints half machine operands correctly.

This cannot currently be hit by any in-tree target.

Patch by Stanislav Mekhanoshin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259857 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix build breakage introduced by r259846.
Easwaran Raman [Fri, 5 Feb 2016 00:45:02 +0000 (00:45 +0000)]
Fix build breakage introduced by r259846.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259855 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd a test for MemorySSA. NFC.
George Burgess IV [Fri, 5 Feb 2016 00:42:02 +0000 (00:42 +0000)]
Add a test for MemorySSA. NFC.

We don't currently have many tests that deal with operations on multiple
local MemoryLocations. This new test helps out a bit in that regard.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259854 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd Support to llvm-c-test dependancies
Amaury Sechet [Fri, 5 Feb 2016 00:19:50 +0000 (00:19 +0000)]
Add Support to llvm-c-test dependancies

Summary: As per title. It is required and don't get linked in in some builds.

Reviewers: chapuni, joker.eph

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D16903

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259853 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFunction name change /NFC
Xinliang David Li [Thu, 4 Feb 2016 23:59:09 +0000 (23:59 +0000)]
Function name change /NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259851 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRefactor profile summary support code. NFC.
Easwaran Raman [Thu, 4 Feb 2016 23:34:31 +0000 (23:34 +0000)]
Refactor profile summary support code. NFC.

Summary computation is not just for instrumented profiling and so I have moved
the ProfileSummary class to ProfileCommon.h (named so to allow code unrelated
to summary but common to instrumented and sampled profiling to be placed there)

Differential Revision: http://reviews.llvm.org/D16661

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259846 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoImprove testing for the C API
Amaury Sechet [Thu, 4 Feb 2016 23:26:19 +0000 (23:26 +0000)]
Improve testing for the C API

Summary:
This basically add an echo test case in C. The support is limited right now, but full support would just be too much to review at once.

The echo test case simply get a module as input and try to output the same exact module. This allow to check the both reading and writing API are working as expected.

I want to improve this test over time to support more and more of the API, in order to improve coverage (coverage is quite poor right now).

Test Plan: Run the test.

Reviewers: chandlerc, bogner

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D10725

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259844 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix for PR 26356
Nemanja Ivanovic [Thu, 4 Feb 2016 23:14:42 +0000 (23:14 +0000)]
Fix for PR 26356

Using the load immediate only when the immediate (whether signed or unsigned)
can fit in a 16-bit signed field. Namely, from -32768 to 32767 for signed and
0 to 65535 for unsigned. This patch also ensures that we sign-extend under the
right conditions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259840 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix typo and test commit
Evandro Menezes [Thu, 4 Feb 2016 23:07:57 +0000 (23:07 +0000)]
Fix typo and test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259839 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoProvide a test case for rl259798
Nemanja Ivanovic [Thu, 4 Feb 2016 22:36:10 +0000 (22:36 +0000)]
Provide a test case for rl259798

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259835 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] Bound the number of instructions we scan when searching for updates.
Chad Rosier [Thu, 4 Feb 2016 21:26:02 +0000 (21:26 +0000)]
[AArch64] Bound the number of instructions we scan when searching for updates.

This only impacts the creation of pre-/post-index instructions.  The bound was
set high enough such that it did not change code generation for SPEC200X.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259828 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[docs] Fix typo in YamlIO.rst
Vedant Kumar [Thu, 4 Feb 2016 20:42:43 +0000 (20:42 +0000)]
[docs] Fix typo in YamlIO.rst

Patch by Mario Lang!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259825 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoInstall cmake files to lib/cmake/llvm
Niels Ole Salscheider [Thu, 4 Feb 2016 20:08:19 +0000 (20:08 +0000)]
Install cmake files to lib/cmake/llvm

This is the right location for platform-specific files.

On some distributions (e. g. Exherbo), a package can be installed for several
architectures in parallel, but the architecture-independent files are shared.
Therefore, we must not install architecture-dependent files (like the CMake
config and export files) to share/.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259821 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][SSE] Select domain for 32/64-bit partial loads for EltsFromConsecutiveLoads
Simon Pilgrim [Thu, 4 Feb 2016 19:27:51 +0000 (19:27 +0000)]
[X86][SSE] Select domain for 32/64-bit partial loads for EltsFromConsecutiveLoads

Choose between MOVD/MOVSS and MOVQ/MOVSD depending on the target vector type.

This has a lot fewer test changes than trying to add this to X86InstrInfo::setExecutionDomain.....

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259816 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix a regression for r259736.
Wei Mi [Thu, 4 Feb 2016 19:17:33 +0000 (19:17 +0000)]
Fix a regression for r259736.

When SCEV expansion tries to reuse an existing value, it is needed to ensure
that using the Value at the InsertPt will not break LCSSA. The fix adds a
check that InsertPt is either inside the candidate Value's parent loop, or
the candidate Value's parent loop is nullptr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259815 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix format in comment
Xinliang David Li [Thu, 4 Feb 2016 19:14:10 +0000 (19:14 +0000)]
Fix format in comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259814 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[PGO] Add interfaces to annotate instr with VP data
Xinliang David Li [Thu, 4 Feb 2016 19:11:43 +0000 (19:11 +0000)]
[PGO] Add interfaces to annotate instr with VP data
  Add interfaces to do value profile data IR annnotation
  and read. Needed by both FE and IR based PGO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259813 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] Improve load/store optimizer to handle LDUR + LDR (take 3).
Chad Rosier [Thu, 4 Feb 2016 18:59:49 +0000 (18:59 +0000)]
[AArch64] Improve load/store optimizer to handle LDUR + LDR (take 3).

This patch allows the mixing of scaled and unscaled load/stores to form
load/store pairs.

PR24465
http://reviews.llvm.org/D12116
Many thanks to Ahmed and Michael for fixes and code review.

This is a reapplication of r246769 and r259790.  The tramp3d failure was caused
by an incorrect refactoring in the patch.  Specifically, we weren't always
properly clearing the SExtIdx flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259812 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SCEV] Add boolean accessors for NSW, NUW and NW; NFC
Sanjoy Das [Thu, 4 Feb 2016 18:21:54 +0000 (18:21 +0000)]
[SCEV] Add boolean accessors for NSW, NUW and NW; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259809 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoCorrectly handle {Always,Never}StepIntoLine
David Majnemer [Thu, 4 Feb 2016 17:57:12 +0000 (17:57 +0000)]
Correctly handle {Always,Never}StepIntoLine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259806 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd support for S_DEFRANGE and S_DEFRANGE_SUBFIELD
David Majnemer [Thu, 4 Feb 2016 17:37:30 +0000 (17:37 +0000)]
Add support for S_DEFRANGE and S_DEFRANGE_SUBFIELD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259805 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMake the dumper's output for variable ranges easier to read
David Majnemer [Thu, 4 Feb 2016 17:29:13 +0000 (17:29 +0000)]
Make the dumper's output for variable ranges easier to read

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259804 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agouse 'auto' for iterators; NFCI
Sanjay Patel [Thu, 4 Feb 2016 17:00:35 +0000 (17:00 +0000)]
use 'auto' for iterators; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259802 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] Multiply extended 32-bit ints with `[U|S]MADDL'
Silviu Baranga [Thu, 4 Feb 2016 16:47:09 +0000 (16:47 +0000)]
[AArch64] Multiply extended 32-bit ints with `[U|S]MADDL'

During instruction selection, the AArch64 backend can recognise the
following pattern and generate an [U|S]MADDL instruction, i.e. a
multiply of two 32-bit operands with a 64-bit result:

(mul (sext i32), (sext i32))
However, when one of the operands is constant, the sign extension
gets folded into the constant in SelectionDAG::getNode(). This means
that the instruction selection sees this:

(mul (sext i32), i64)
...which doesn't match the pattern. Sign-extension and 64-bit
multiply instructions are generated, which are slower than one 32-bit
multiply.

Add a pattern to match this and generate the correct instruction, for
both signed and unsigned multiplies.

Patch by Chris Diamand!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259800 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoThe canonical way to XFAIL a test for all targets is XFAIL: *, not XFAIL:
Benjamin Kramer [Thu, 4 Feb 2016 16:21:38 +0000 (16:21 +0000)]
The canonical way to XFAIL a test for all targets is XFAIL: *, not XFAIL:

Fix the lit bug that enabled this "feature" (empty triple is substring
of all possible target triples) and change the two outliers to use the
documented * syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259799 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoEnable the %s modifier in inline asm template string
Nemanja Ivanovic [Thu, 4 Feb 2016 16:18:08 +0000 (16:18 +0000)]
Enable the %s modifier in inline asm template string

This patch corresponds to review:
http://reviews.llvm.org/D16847

There are some files in glibc that use the output operand modifier even though
it was deprecated in GCC. This patch just adds support for it to prevent issues
with such files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259798 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[PPC] Move PPC test to a PPC-specific dir
Renato Golin [Thu, 4 Feb 2016 16:14:59 +0000 (16:14 +0000)]
[PPC] Move PPC test to a PPC-specific dir

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259797 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][SSE] Add general 32-bit LOAD + VZEXT_MOVL support to EltsFromConsecutiveLoads
Simon Pilgrim [Thu, 4 Feb 2016 16:12:56 +0000 (16:12 +0000)]
[X86][SSE] Add general 32-bit LOAD + VZEXT_MOVL support to EltsFromConsecutiveLoads

This patch adds support for consecutive (load/undef elements) 32-bit loads, followed by trailing undef/zero elements to be combined to a single MOVD load.

Differential Revision: http://reviews.llvm.org/D16729

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259796 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "[AArch64] Improve load/store optimizer to handle LDUR + LDR."
Chad Rosier [Thu, 4 Feb 2016 16:01:40 +0000 (16:01 +0000)]
Revert "[AArch64] Improve load/store optimizer to handle LDUR + LDR."

This reverts commit r259790. tramp3d-v4 is still having problems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259795 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][SSE] Added i686 target tests to make sure we are correctly loading consecutive...
Simon Pilgrim [Thu, 4 Feb 2016 15:51:55 +0000 (15:51 +0000)]
[X86][SSE] Added i686 target tests to make sure we are correctly loading consecutive entries as 64-bit integers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259794 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAVX-512: Fixed a bug in FMA instruction selection on KNL
Elena Demikhovsky [Thu, 4 Feb 2016 15:11:11 +0000 (15:11 +0000)]
AVX-512: Fixed a bug in FMA instruction selection on KNL

The FMA instruction was selected from AVX2 set instead of AVX-512

Differential Revision: http://reviews.llvm.org/D16884

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259792 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[Power PC] softening long double type
Petar Jovanovic [Thu, 4 Feb 2016 14:43:50 +0000 (14:43 +0000)]
[Power PC] softening long double type

This patch implements softening of long double type (ppcf128) on ppc32
architecture and enables operations for this type for soft float.

Patch by Strahinja Petrovic.

Differential Revision: http://reviews.llvm.org/D15811

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259791 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] Improve load/store optimizer to handle LDUR + LDR.
Chad Rosier [Thu, 4 Feb 2016 14:42:55 +0000 (14:42 +0000)]
[AArch64] Improve load/store optimizer to handle LDUR + LDR.

This patch allows the mixing of scaled and unscaled load/stores to form
load/store pairs.

PR24465
http://reviews.llvm.org/D12116
Many thanks to Ahmed and Michael for fixes and code review.

This is a reapplication of r246769, which was reverted in r246782 due to a
test-suite failure.  I'm unable to reproduce the issue at this time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259790 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AVX512] add vfmadd132ss and vfmadd132sd Intrinsic
Michael Zuckerman [Thu, 4 Feb 2016 14:41:08 +0000 (14:41 +0000)]
[AVX512] add vfmadd132ss and vfmadd132sd Intrinsic

Differential Revision: http://reviews.llvm.org/D16589

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259789 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Add AVX512 vector zext tests
Simon Pilgrim [Thu, 4 Feb 2016 14:06:19 +0000 (14:06 +0000)]
[X86] Add AVX512 vector zext tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259786 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ScheduleDagInstrs] Improved comments
Jonas Paulsson [Thu, 4 Feb 2016 13:08:48 +0000 (13:08 +0000)]
[ScheduleDagInstrs] Improved comments

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259783 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Moved SEXT -> SIGN_EXTEND_VECTOR_INREG combine into helper. NFC.
Simon Pilgrim [Thu, 4 Feb 2016 09:27:19 +0000 (09:27 +0000)]
[X86] Moved SEXT -> SIGN_EXTEND_VECTOR_INREG combine into helper. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259771 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Use hash table in LEA optimization pass.
Andrey Turetskiy [Thu, 4 Feb 2016 08:57:03 +0000 (08:57 +0000)]
[X86] Use hash table in LEA optimization pass.

Use hash table (key is a memory operand) to store found LEA instructions to reduce compile time.

Differential Revision: http://reviews.llvm.org/D16404

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259770 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agocmake: Add a flag to enable LTO
Justin Bogner [Thu, 4 Feb 2016 07:28:30 +0000 (07:28 +0000)]
cmake: Add a flag to enable LTO

This adds -DLLVM_ENABLE_LTO, rather than forcing people to manually
add -flto to the various _FLAGS variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259766 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[Support] Use range-based for loop. NFC
Craig Topper [Thu, 4 Feb 2016 06:51:41 +0000 (06:51 +0000)]
[Support] Use range-based for loop. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259763 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[Support] Use hexdigit instead of manually coding the same thing. NFC
Craig Topper [Thu, 4 Feb 2016 06:51:38 +0000 (06:51 +0000)]
[Support] Use hexdigit instead of manually coding the same thing. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259762 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[PGO] Profile interface cleanup
Xinliang David Li [Thu, 4 Feb 2016 05:29:51 +0000 (05:29 +0000)]
[PGO] Profile interface cleanup
  - Remove unused valuemapper parameter
  - add totalcount optional parameter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259756 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[NVPTX] Disable performance optimizations when OptLevel==None
Jingyue Wu [Thu, 4 Feb 2016 04:15:36 +0000 (04:15 +0000)]
[NVPTX] Disable performance optimizations when OptLevel==None

Reviewers: jholewinski, tra, eliben

Subscribers: jholewinski, llvm-commits

Differential Revision: http://reviews.llvm.org/D16874

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259749 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoTest case for PR 26381
Nemanja Ivanovic [Thu, 4 Feb 2016 01:58:20 +0000 (01:58 +0000)]
Test case for PR 26381

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259740 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SCEV] Try to reuse existing value during SCEV expansion
Wei Mi [Thu, 4 Feb 2016 01:27:38 +0000 (01:27 +0000)]
[SCEV] Try to reuse existing value during SCEV expansion

Current SCEV expansion will expand SCEV as a sequence of operations
and doesn't utilize the value already existed. This will introduce
redundent computation which may not be cleaned up throughly by
following optimizations.

This patch introduces an ExprValueMap which is a map from SCEV to the
set of equal values with the same SCEV. When a SCEV is expanded, the
set of values is checked and reused whenever possible before generating
a sequence of operations.

The original commit triggered regressions in Polly tests. The regressions
exposed two problems which have been fixed in current version.

1. Polly will generate a new function based on the old one. To generate an
instruction for the new function, it builds SCEV for the old instruction,
applies some tranformation on the SCEV generated, then expands the transformed
SCEV and insert the expanded value into new function. Because SCEV expansion
may reuse value cached in ExprValueMap, the value in old function may be
inserted into new function, which is wrong.
   In SCEVExpander::expand, there is a logic to check the cached value to
be used should dominate the insertion point. However, for the above
case, the check always passes. That is because the insertion point is
in a new function, which is unreachable from the old function. However
for unreachable node, DominatorTreeBase::dominates thinks it will be
dominated by any other node.
   The fix is to simply add a check that the cached value to be used in
expansion should be in the same function as the insertion point instruction.

2. When the SCEV is of scConstant type, expanding it directly is cheaper than
reusing a normal value cached. Although in the cached value set in ExprValueMap,
there is a Constant type value, but it is not easy to find it out -- the cached
Value set is not sorted according to the potential cost. Existing reuse logic
in SCEVExpander::expand simply chooses the first legal element from the cached
value set.
   The fix is that when the SCEV is of scConstant type, don't try the reuse
logic. simply expand it.

Differential Revision: http://reviews.llvm.org/D12090

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259736 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix undefined behavior when compiling in C++14 mode (with sized deletion
Richard Smith [Thu, 4 Feb 2016 01:21:16 +0000 (01:21 +0000)]
Fix undefined behavior when compiling in C++14 mode (with sized deletion
enabled): ensure that we do not invoke the sized deallocator for MemoryBuffer
subclasses that have tail-allocated data.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259735 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[codeview] Don't attempt a cross-section label diff
Reid Kleckner [Thu, 4 Feb 2016 00:21:42 +0000 (00:21 +0000)]
[codeview] Don't attempt a cross-section label diff

This only comes up when we're trying to find the next .cv_loc label.

Fixes PR26467

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259733 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[libFuzzer] hot fix a test
Kostya Serebryany [Thu, 4 Feb 2016 00:12:28 +0000 (00:12 +0000)]
[libFuzzer] hot fix a test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259732 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[libFuzzer] don't write the test unit when a leak is detected (since we don't know...
Kostya Serebryany [Thu, 4 Feb 2016 00:02:17 +0000 (00:02 +0000)]
[libFuzzer] don't write the test unit when a leak is detected (since we don't know which unit causes the leak)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259731 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SimplifyCFG] Fix for "endless" loop after dead code removal (Alternative to
Gerolf Hoflehner [Wed, 3 Feb 2016 23:54:25 +0000 (23:54 +0000)]
[SimplifyCFG] Fix for "endless" loop after dead code removal (Alternative to
D16251)

Summary:
This is a simpler fix to the problem than the dominator approach in
http://reviews.llvm.org/D16251. It adds only values into the gather() while loop
that have been seen before.

The actual endless loop is in the constant compare gather() routine in
Utils/SimplifyCFG.cpp. The same value ret.0.off0.i is pushed back into the
queue:
%.ret.0.off0.i = or i1 %.ret.0.off0.i, %cmp10.i

Here is what happens at the IR level:

for.cond.i:                                       ; preds = %if.end6.i,
%if.end.i54
%ix.0.i = phi i32 [ 0, %if.end.i54 ], [ %inc.i55, %if.end6.i ]
%ret.0.off0.i = phi i1 [false, %if.end.i54], [%.ret.0.off0.i, %if.end6.i] <<<
%cmp2.i = icmp ult i32 %ix.0.i, %11
br i1 %cmp2.i, label %for.body.i, label %LBJ_TmpSimpleNeedExt.exit

if.end6.i:                                        ; preds = %for.body.i
%cmp10.i = icmp ugt i32 %conv.i, %add9.i
%.ret.0.off0.i = or i1 %ret.0.off0.i, %cmp10.i <<<

When if.end.i54 gets eliminated which removes the definition of ret.0.off0.i.
The result is the expression %.ret.0.off0.i = or i1 %.ret.0.off0.i, %cmp10.i
(Note the first ‘or’ operand is now %.ret.0.off0.i, and *NOT* %ret.0.off0.i).
And
now there is use of .ret.0.off0.i before a definition which triggers the
“endless” loop in gather():

while(!DFT.empty()) {

    V = DFT.pop_back_val();   // V is .ret.0.off0.i

    if (Instruction *I = dyn_cast<Instruction>(V)) {
      // If it is a || (or && depending on isEQ), process the operands.
      if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
        DFT.push_back(I->getOperand(1));  // This is now .ret.0.off0.i also
        DFT.push_back(I->getOperand(0));

        continue; // “endless loop” for .ret.0.off0.i
      }

Reviewers: reames, ahatanak

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D16839

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259730 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstrProfiling] Fix a comment (NFC)
Vedant Kumar [Wed, 3 Feb 2016 23:22:43 +0000 (23:22 +0000)]
[InstrProfiling] Fix a comment (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259727 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoUnify the target opcode enum in TargetOpcodes.h and the FixedInstrs array in
David L Kreitzer [Wed, 3 Feb 2016 23:17:32 +0000 (23:17 +0000)]
Unify the target opcode enum in TargetOpcodes.h and the FixedInstrs array in
CodeGenTarget.cpp to avoid the ordering dependence. NFCI.

Differential Revision: http://reviews.llvm.org/D16826

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259726 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMinor code cleanups. NFC.
Junmo Park [Wed, 3 Feb 2016 23:16:39 +0000 (23:16 +0000)]
Minor code cleanups. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259725 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoPrint the OffsetStart field's relocation
David Majnemer [Wed, 3 Feb 2016 22:45:21 +0000 (22:45 +0000)]
Print the OffsetStart field's relocation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259723 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agorangify; NFCI
Sanjay Patel [Wed, 3 Feb 2016 22:44:14 +0000 (22:44 +0000)]
rangify; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259722 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoclean up; NFC
Sanjay Patel [Wed, 3 Feb 2016 22:37:37 +0000 (22:37 +0000)]
clean up; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259720 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[llvm-readobj] Add support for dumping S_DEFRANGE symbols
David Majnemer [Wed, 3 Feb 2016 22:36:46 +0000 (22:36 +0000)]
[llvm-readobj] Add support for dumping S_DEFRANGE symbols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259719 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoReplace static const int with enum to fix obnoxious linker errors about a missing...
Reid Kleckner [Wed, 3 Feb 2016 21:45:39 +0000 (21:45 +0000)]
Replace static const int with enum to fix obnoxious linker errors about a missing definition

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259712 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[unittests] Move TargetRegistry test from Support to MC
Reid Kleckner [Wed, 3 Feb 2016 21:41:24 +0000 (21:41 +0000)]
[unittests] Move TargetRegistry test from Support to MC

This removes the dependency from SupportTests to all of the LLVM
backends, and makes it link faster.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259705 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoSilence -Wsign-conversion issue in ProgramTest.cpp
Reid Kleckner [Wed, 3 Feb 2016 21:41:12 +0000 (21:41 +0000)]
Silence -Wsign-conversion issue in ProgramTest.cpp

Unfortunately, ProgramInfo::ProcessId is signed on Unix and unsigned on
Windows, breaking the standard fix of using '0U' in the gtest
expectation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259704 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix pointers to go on the right hand side. NFC.
Ana Pazos [Wed, 3 Feb 2016 21:34:39 +0000 (21:34 +0000)]
Fix pointers to go on the right hand side. NFC.

Summary:
Fixed pointers to go on the right hand side following coding guidelines. NFC.

Patch by Mandeep Singh Grang.

Reviewers: majnemer, arsenm, sanjoy

Differential Revision: http://reviews.llvm.org/D16866

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259703 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LoopStrengthReduce] Don't rewrite PHIs with incoming values from CatchSwitches
David Majnemer [Wed, 3 Feb 2016 21:30:34 +0000 (21:30 +0000)]
[LoopStrengthReduce] Don't rewrite PHIs with incoming values from CatchSwitches

Bail out if we have a PHI on an EHPad that gets a value from a
CatchSwitchInst.  Because the CatchSwitchInst cannot be split, there is
no good place to stick any instructions.

This fixes PR26373.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259702 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ScalarEvolutionExpander] Simplify findInsertPointAfter
David Majnemer [Wed, 3 Feb 2016 21:30:31 +0000 (21:30 +0000)]
[ScalarEvolutionExpander] Simplify findInsertPointAfter

No functional change is intended.  The loop could only execute, at most,
once.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259701 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[codeview] Remove EmitLabelDiff in favor emitAbsoluteSymbolDiff
Reid Kleckner [Wed, 3 Feb 2016 21:24:42 +0000 (21:24 +0000)]
[codeview] Remove EmitLabelDiff in favor emitAbsoluteSymbolDiff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259700 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[codeview] Use the MCStreamer interface directly instead of AsmPrinter
Reid Kleckner [Wed, 3 Feb 2016 21:15:48 +0000 (21:15 +0000)]
[codeview] Use the MCStreamer interface directly instead of AsmPrinter

This is mostly about having shorter lines and standardizing on one
interface, but it also avoids some needless indirection.

No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259697 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[DWARFDebug] Fix another case of overlapping ranges
Keno Fischer [Wed, 3 Feb 2016 21:13:33 +0000 (21:13 +0000)]
[DWARFDebug] Fix another case of overlapping ranges

Summary:
In r257979, I added code to ensure that we wouldn't merge DebugLocEntries if
the pieces they describe overlap. Unfortunately, I failed to cover the case,
where there may have multiple active Expressions in the entry, in which case we
need to make sure that no two values overlap before we can perform the merge.

This fixed PR26148.

Reviewers: aprantl
Differential Revision: http://reviews.llvm.org/D16742

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259696 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAddress NDEBUG-related linkage issues for Value::assertModuleIsMaterialized()
Todd Fiala [Wed, 3 Feb 2016 21:13:23 +0000 (21:13 +0000)]
Address NDEBUG-related linkage issues for Value::assertModuleIsMaterialized()

The IR/Value class had a linkage issue present when LLVM was built
as a library, and the LLVM library build time had different settings
for NDEBUG than the client of the LLVM library.  Clients could get
into a state where the LLVM lib expected
Value::assertModuleIsMaterialized() to be inline-defined in the header
but clients expected that method to be defined in the LLVM library.

See this llvm-commits thread for more details:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160201/329667.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259695 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SelectionDAG] Fix CombineToPreIndexedLoadStore O(n^2) behavior
Tim Shen [Wed, 3 Feb 2016 20:58:55 +0000 (20:58 +0000)]
[SelectionDAG] Fix CombineToPreIndexedLoadStore O(n^2) behavior

This patch consists of two parts: a performance fix in DAGCombiner.cpp
and a correctness fix in SelectionDAG.cpp.

The test case tests the bug that's uncovered by the performance fix, and
fixed by the correctness fix.

The performance fix keeps the containers required by the
hasPredecessorHelper (which is a lazy DFS) and reuse them. Since
hasPredecessorHelper is called in a loop, the overall efficiency reduced
from O(n^2) to O(n), where n is the number of SDNodes.

The correctness fix keeps iterating the neighbor list even if it's time
to early return. It will return after finishing adding all neighbors to
Worklist, so that no neighbors are discarded due to the original early
return.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259691 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMinor performance tweaks to llvm-tblgen (and a few that might be a good idea)
Reid Kleckner [Wed, 3 Feb 2016 19:34:28 +0000 (19:34 +0000)]
Minor performance tweaks to llvm-tblgen (and a few that might be a good idea)

Summary:
This patch adds a reserve call to an expensive function
(`llvm::LoadIntrinsics`), and may fix a few other low hanging
performance fruit (I've put them in comments for now, so we can
discuss).

**Motivation:**

As I'm sure other developers do, when I build LLVM, I build the entire
project with the same config (`Debug`, `MinSizeRel`, `Release`, or
`RelWithDebInfo`). However, the `Debug` config also builds llvm-tblgen
in `Debug` mode. Later build steps that run llvm-tblgen then can
actually be the slowest steps in the entire build. Nobody likes slow
builds.

Reviewers: rnk, dblaikie

Differential Revision: http://reviews.llvm.org/D16832

Patch by Alexander G. Riccio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259683 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoARM: support TLS for WoA
Saleem Abdulrasool [Wed, 3 Feb 2016 18:21:59 +0000 (18:21 +0000)]
ARM: support TLS for WoA

Add support for TLS access for Windows on ARM.  This generates a similar access
to MSVC for ARM.

The changes to the tablegen data is needed to support loading an external symbol
global that is not for a call.  The adjustments to the DAG to DAG transforms are
needed to preserve the 32-bit move.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259676 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert r259662, which caused regressions on polly tests.
Wei Mi [Wed, 3 Feb 2016 18:05:57 +0000 (18:05 +0000)]
Revert r259662, which caused regressions on polly tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259675 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstCombine] Revert r238452: Fold IntToPtr and PtrToInt into preceding loads.
Quentin Colombet [Wed, 3 Feb 2016 18:04:13 +0000 (18:04 +0000)]
[InstCombine] Revert r238452: Fold IntToPtr and PtrToInt into preceding loads.

According to git bisect, this is the root cause of a miscompile for Regex in
libLLVMSupport. I am still working on reducing a test case.
The actual bug may be elsewhere and this commit just exposed it.

Anyway, at the moment, to reproduce, follow these steps:
1. Build clang and libLTO in release mode.
2. Create a new build directory <stage2> and cd into it.
3. Use clang and libLTO from #1 to build llvm-extract in Release mode + asserts
   using -O2 -flto
4. Run llvm-extract  -ralias '.*bar' -S test/Other/extract-alias.ll

Result:
program doesn't contain global named '.*bar'!

Expected result:
@a0a0bar = alias void ()* @bar
@a0bar = alias void ()* @bar

declare void @bar()

Note: In step #3, if you don't use lto or asserts, the miscompile disappears.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259674 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ScheduleDAGInstrs::buildSchedGraph()] Handling of memory dependecies rewritten.
Jonas Paulsson [Wed, 3 Feb 2016 17:52:29 +0000 (17:52 +0000)]
[ScheduleDAGInstrs::buildSchedGraph()] Handling of memory dependecies rewritten.

Recommited, after some fixing with test cases.

Updated test cases:
test/CodeGen/AArch64/arm64-misched-memdep-bug.ll
test/CodeGen/AArch64/tailcall_misched_graph.ll

Temporarily disabled test cases:
test/CodeGen/AMDGPU/split-vector-memoperand-offsets.ll
test/CodeGen/PowerPC/ppc64-fastcc.ll (partially updated)
test/CodeGen/PowerPC/vsx-fma-m.ll
test/CodeGen/PowerPC/vsx-fma-sp.ll

http://reviews.llvm.org/D8705
Reviewers: Hal Finkel, Andy Trick.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259673 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix comments /NFC
Xinliang David Li [Wed, 3 Feb 2016 17:51:16 +0000 (17:51 +0000)]
Fix comments /NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259672 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[Unittest] Clean up formatting, NFC
Joseph Tremoulet [Wed, 3 Feb 2016 17:11:24 +0000 (17:11 +0000)]
[Unittest] Clean up formatting, NFC

Summary:
Use an early return to reduce indentation.
Remove unused local.

Reviewers: dblaikie, lhames

Subscribers: lhames, llvm-commits

Differential Revision: http://reviews.llvm.org/D16513

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259663 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SCEV] Try to reuse existing value during SCEV expansion
Wei Mi [Wed, 3 Feb 2016 17:05:12 +0000 (17:05 +0000)]
[SCEV] Try to reuse existing value during SCEV expansion

Current SCEV expansion will expand SCEV as a sequence of operations
and doesn't utilize the value already existed. This will introduce
redundent computation which may not be cleaned up throughly by
following optimizations.

This patch introduces an ExprValueMap which is a map from SCEV to the
set of equal values with the same SCEV. When a SCEV is expanded, the
set of values is checked and reused whenever possible before generating
a sequence of operations.

Differential Revision: http://reviews.llvm.org/D12090

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259662 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ARM] Move GNUEABI divmod to __aeabi_divmod*
Renato Golin [Wed, 3 Feb 2016 16:10:54 +0000 (16:10 +0000)]
[ARM] Move GNUEABI divmod to __aeabi_divmod*

The GNU toolchain emits __aeabi_divmod for soft-divide on ARM cores
which happens to be a lot faster than __divsi3/__modsi3 when the core
has hardware divide instructions. Do the same here.

Fixes PR26450.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259657 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[MachineCopyPropagation] Fix comment. NFC
Jun Bum Lim [Wed, 3 Feb 2016 15:56:27 +0000 (15:56 +0000)]
[MachineCopyPropagation] Fix comment. NFC

Reviewers: MatzeB, qcolombet, jmolloy, mcrosier

Subscribers: llvm-commits, mcrosier

Differential Revision: http://reviews.llvm.org/D16806

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259656 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[mips] Remove redundant inclusions of MipsAnalyzeImmediate.h
Daniel Sanders [Wed, 3 Feb 2016 15:54:12 +0000 (15:54 +0000)]
[mips] Remove redundant inclusions of MipsAnalyzeImmediate.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259655 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[DemandedBits] Revert r249687 due to PR26071
James Molloy [Wed, 3 Feb 2016 15:05:06 +0000 (15:05 +0000)]
[DemandedBits] Revert r249687 due to PR26071

This regresses a test in LoopVectorize, so I'll need to go away and think about how to solve this in a way that isn't broken.

From the writeup in PR26071:

What's happening is that ComputeKnownZeroes is telling us that all bits except the LSB are zero. We're then deciding that only the LSB needs to be demanded from the icmp's inputs.

This is where we're wrong - we're assuming that after simplification the bits that were known zero will continue to be known zero. But they're not - during trivialization the upper bits get changed (because an XOR isn't shrunk), so the icmp fails.

The fault is in demandedbits - its contract does clearly state that a non-demanded bit may either be zero or one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259649 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix for PR 26381
Nemanja Ivanovic [Wed, 3 Feb 2016 12:53:38 +0000 (12:53 +0000)]
Fix for PR 26381

Simple fix - Constant values were not being sign extended in FastIsel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259645 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[mips] Add SHF_MIPS_GPREL flag to the MIPS .sbss and .sdata sections
Simon Atanasyan [Wed, 3 Feb 2016 11:50:22 +0000 (11:50 +0000)]
[mips] Add SHF_MIPS_GPREL flag to the MIPS .sbss and .sdata sections

MIPS ABI states that .sbss and .sdata sections must have SHF_MIPS_GPREL
flag. See Figure 4–7 on page 69 in the following document:
ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf.

Differential Revision: http://reviews.llvm.org/D15740

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259641 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[TableGen] Add 'register alternative name matching' support
Dylan McKay [Wed, 3 Feb 2016 10:30:16 +0000 (10:30 +0000)]
[TableGen] Add 'register alternative name matching' support

Summary:
This adds a new attribute which targets can set in TableGen which causes a function to be generated which matches register alternative names. This is very similar to `ShouldEmitMatchRegisterName`, except it works on alt names.

This patch is currently used by the out of tree part of the AVR backend. It reduces code duplication greatly, and has the effect that you do not need to hardcode altname to register mappings in C++.

It will not work on targets which have registers which share the same aliases.

Reviewers: stoklund, arsenm, dsanders, hfinkel, vkalintiris

Subscribers: hfinkel, dylanmckay, llvm-commits

Differential Revision: http://reviews.llvm.org/D16312

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259636 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][AVX] Add support for 64-bit VZEXT_LOAD of 256/512-bit vectors to EltsFromConsec...
Simon Pilgrim [Wed, 3 Feb 2016 09:41:59 +0000 (09:41 +0000)]
[X86][AVX] Add support for 64-bit VZEXT_LOAD of 256/512-bit vectors to EltsFromConsecutiveLoads

Follow up to D16217 and D16729

This change uncovered an odd pattern where VZEXT_LOAD v4i64 was being lowered to a load of the lower v2i64 (so the 2nd i64 destination element wasn't being zeroed), I can't find any use/reason for this and have removed the pattern and replaced it so only the 1st i64 element is loaded and the upper bits all zeroed. This matches the description for X86ISD::VZEXT_LOAD

Differential Revision: http://reviews.llvm.org/D16768

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259635 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd a compatibility test
Xinliang David Li [Wed, 3 Feb 2016 06:27:38 +0000 (06:27 +0000)]
Add a compatibility test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259632 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix a typo in comment
Xinliang David Li [Wed, 3 Feb 2016 06:24:11 +0000 (06:24 +0000)]
Fix a typo in comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@259631 91177308-0d34-0410-b5e6-96231b3b80d8