OSDN Git Service

android-x86/external-llvm.git
7 years agoAdd missing library dep.
Peter Collingbourne [Fri, 16 Dec 2016 00:43:00 +0000 (00:43 +0000)]
Add missing library dep.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289903 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[IR] Remove the DIExpression field from DIGlobalVariable.
Adrian Prantl [Fri, 16 Dec 2016 00:36:43 +0000 (00:36 +0000)]
[IR] Remove the DIExpression field from DIGlobalVariable.

This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.

Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:

(1) The DIGlobalVariable should describe the source level variable,
    not how to get to its location.

(2) It makes it unsafe/hard to update the expressions when we call
    replaceExpression on the DIGLobalVariable.

(3) It makes it impossible to represent a global variable that is in
    more than one location (e.g., a variable with multiple
    DW_OP_LLVM_fragment-s).  We also moved away from attaching the
    DIExpression to DILocalVariable for the same reasons.

<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289902 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PPC] corrections in two testcases
Ehsan Amiri [Fri, 16 Dec 2016 00:33:07 +0000 (00:33 +0000)]
[PPC] corrections in two testcases

Removing sensitivity to scheduling (by using CHECK-DAG instead of CHECK) and
some other minor corrections.

In preparation to commit Power9 processor model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289900 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIPO: Introduce ThinLTOBitcodeWriter pass.
Peter Collingbourne [Fri, 16 Dec 2016 00:26:30 +0000 (00:26 +0000)]
IPO: Introduce ThinLTOBitcodeWriter pass.

This pass prepares a module containing type metadata for ThinLTO by splitting
it into regular and thin LTO parts if possible, and writing both parts to
a multi-module bitcode file. Modules that do not contain type metadata are
written unmodified as a single module.

All globals with type metadata are added to the regular LTO module, and
the rest are added to the thin LTO module.

Differential Revision: https://reviews.llvm.org/D27324

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289899 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Add FeatureSlowMisaligned128Store to Exynos M1 and M2
Evandro Menezes [Fri, 16 Dec 2016 00:18:00 +0000 (00:18 +0000)]
[AArch64] Add FeatureSlowMisaligned128Store to Exynos M1 and M2

This feature now gates such stores after r289845.  Thus the Exynos
processors now need this feature.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289898 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Thin link efficiency improvement: don't re-export globals (NFC)
Teresa Johnson [Thu, 15 Dec 2016 23:50:06 +0000 (23:50 +0000)]
[ThinLTO] Thin link efficiency improvement: don't re-export globals (NFC)

Summary:
We were reinvoking exportGlobalInModule numerous times redundantly.
No need to re-export globals referenced by a global that was already
imported from its module. This resulted in a large speedup in the thin
link for a big application, particularly when importing aggressiveness
was cranked up.

Reviewers: mehdi_amini

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27687

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289896 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SimplifyLibCalls] Add a test to make sure we lower fls(0) correctly.
Davide Italiano [Thu, 15 Dec 2016 23:48:07 +0000 (23:48 +0000)]
[SimplifyLibCalls] Add a test to make sure we lower fls(0) correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289895 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SimplifyLibCalls] Lower fls() to llvm.ctlz().
Davide Italiano [Thu, 15 Dec 2016 23:45:11 +0000 (23:45 +0000)]
[SimplifyLibCalls] Lower fls() to llvm.ctlz().

Differential Revision:  https://reviews.llvm.org/D14590

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289894 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDebugInfo: Make a Generic test case actually generic (remove datalayout/triple)
David Blaikie [Thu, 15 Dec 2016 23:39:25 +0000 (23:39 +0000)]
DebugInfo: Make a Generic test case actually generic (remove datalayout/triple)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289893 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDebugInfo: Address non-deterministic output (iterating a SmallPtrSet) in 289697
David Blaikie [Thu, 15 Dec 2016 23:37:38 +0000 (23:37 +0000)]
DebugInfo: Address non-deterministic output (iterating a SmallPtrSet) in 289697

Post-commit review feedback from Adrian Prantl.

Hopefully this fixes that up :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289892 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[IRTranslator] Merge the entry and ABI lowering blocks.
Quentin Colombet [Thu, 15 Dec 2016 23:32:25 +0000 (23:32 +0000)]
[IRTranslator] Merge the entry and ABI lowering blocks.

The IRTranslator uses an additional block before the LLVM-IR entry block
to perform all the ABI lowering and the constant hoisting. Thus, this
block is the actual entry block and it falls through the LLVM-IR entry
block. However, with such representation, we end up with two basic
blocks that are not maximal.

Therefore, this patch adds a bit of canonicalization by merging both the
LLVM-IR entry block and the ABI lowering/constants hoisting into one
block, making the resulting block more likely to be maximal (indeed the
LLVM-IR entry block might not have been maximal).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289891 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDebugInfo: Emit ranges for functions with DISubprograms but lacking locations on...
David Blaikie [Thu, 15 Dec 2016 23:17:52 +0000 (23:17 +0000)]
DebugInfo: Emit ranges for functions with DISubprograms but lacking locations on any instructions

This seems more consistent, and helps tidy up/simplify some other code
in this change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289889 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SimplifyLibCalls] Remove redundant folding logic for ffs().
Davide Italiano [Thu, 15 Dec 2016 23:11:00 +0000 (23:11 +0000)]
[SimplifyLibCalls] Remove redundant folding logic for ffs().

Lowering to llvm.cttz() will result in constant folding anyway
if the argument to ffs is a constant. Pointed out by Eli for
fls() in D14590.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289888 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDon't combine splats with other shuffles.
Eli Friedman [Thu, 15 Dec 2016 22:41:40 +0000 (22:41 +0000)]
Don't combine splats with other shuffles.

We sometimes end up creating shuffles which are worse than the obvious
translation of the IR.

Fixes https://llvm.org/bugs/show_bug.cgi?id=31301 .

Differential Revision: https://reviews.llvm.org/D27793

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289882 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix R_AARCH64_MOVW_UABS_G3 relocation
Yichao Yu [Thu, 15 Dec 2016 22:36:53 +0000 (22:36 +0000)]
Fix R_AARCH64_MOVW_UABS_G3 relocation

Summary: The relocation is missing mask so an address that has non-zero bits in 47:43 may overwrite the register number. (Frequently shows up as target register changed to `xzr`....)

Reviewers: t.p.northover, lhames

Subscribers: davide, aemerson, rengolin, llvm-commits

Differential Revision: https://reviews.llvm.org/D27609

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289880 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Select branch on undef to uniform scc branch
Matt Arsenault [Thu, 15 Dec 2016 21:57:11 +0000 (21:57 +0000)]
AMDGPU: Select branch on undef to uniform scc branch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289877 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[gold] Add datalayout to test where it was missing
Teresa Johnson [Thu, 15 Dec 2016 21:42:56 +0000 (21:42 +0000)]
[gold] Add datalayout to test where it was missing

Needed due to change to require datalayout (r289719).

Found this in my own testing, maybe there aren't any bots using a v1.12
gold yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289876 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Revert part of r289843 that belonged to another patch.
Teresa Johnson [Thu, 15 Dec 2016 21:39:42 +0000 (21:39 +0000)]
[ThinLTO] Revert part of r289843 that belonged to another patch.

The code change for D27687 accidentally got committed along with the
main change in r289843. Revert it temporarily, so that I can recommit it
along with its test as intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289875 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDon't combine a shuffle of two BUILD_VECTORs with duplicate elements.
Eli Friedman [Thu, 15 Dec 2016 21:36:59 +0000 (21:36 +0000)]
Don't combine a shuffle of two BUILD_VECTORs with duplicate elements.

Targets can't handle this case well in general; we often transform
a shuffle of two cheap BUILD_VECTORs to element-by-element insertion,
which is very inefficient.

Fixes https://llvm.org/bugs/show_bug.cgi?id=31364 . Partially
fixes https://llvm.org/bugs/show_bug.cgi?id=31301.

Differential Revision: https://reviews.llvm.org/D27787

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289874 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Verifier] Allow TBAA metadata on atomicrmw and atomiccmpxchg
Sanjoy Das [Thu, 15 Dec 2016 21:23:44 +0000 (21:23 +0000)]
[Verifier] Allow TBAA metadata on atomicrmw and atomiccmpxchg

This used to be allowed before r289402 by default (before r289402 you
could have TBAA metadata on any instruction), and while I'm not sure
that it helps, it does sound reasonable enough to not fail the verifier
and we have out-of-tree users who use this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289872 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Remove stale comment (NFC)
Teresa Johnson [Thu, 15 Dec 2016 20:53:31 +0000 (20:53 +0000)]
[ThinLTO] Remove stale comment (NFC)

This should have been removed with r288446.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289871 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PPC] Use CHECK-DAG instead of CHECK in the testcase
Ehsan Amiri [Thu, 15 Dec 2016 20:51:09 +0000 (20:51 +0000)]
[PPC] Use CHECK-DAG instead of CHECK in the testcase

This test is currently sensitive to scheduling. Using CHECK-DAG allows us to
preserve the main purpose of the test and remove this sensivity.

In preparation to commit Power9 processor model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289869 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Fix asserting on returned tail calls
Matt Arsenault [Thu, 15 Dec 2016 20:50:12 +0000 (20:50 +0000)]
AMDGPU: Fix asserting on returned tail calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289868 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Thin link efficiency: skip candidate added later with higher threshold...
Teresa Johnson [Thu, 15 Dec 2016 20:48:19 +0000 (20:48 +0000)]
[ThinLTO] Thin link efficiency: skip candidate added later with higher threshold (NFC)

Summary:
Thin link efficiency improvement. After adding an importing candidate to
the worklist we might have later added it again with a higher threshold.
Skip it when popped from the worklist if we recorded a higher threshold
than the current worklist entry, it will get processed again at the
higher threshold when that entry is popped.

This required adding the summary's GUID to the worklist, so that it can
be used to query the recorded highest threshold for it when we pop from the
worklist.

Reviewers: mehdi_amini

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27696

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289867 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Assembler support for vintrp instructions
Matt Arsenault [Thu, 15 Dec 2016 20:40:20 +0000 (20:40 +0000)]
AMDGPU: Assembler support for vintrp instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289866 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LV] Enable vectorization of loops with conditional stores by default
Matthew Simpson [Thu, 15 Dec 2016 20:11:05 +0000 (20:11 +0000)]
[LV] Enable vectorization of loops with conditional stores by default

This patch sets the default value of the "-enable-cond-stores-vec" command line
option to "true".

Differential Revision: https://reviews.llvm.org/D27814

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289863 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SimplifyCFG] Merge debug locations when hoisting an instruction from a then/else...
Andrea Di Biagio [Thu, 15 Dec 2016 20:01:26 +0000 (20:01 +0000)]
[SimplifyCFG] Merge debug locations when hoisting an instruction from a then/else branch. NFC.

Now that a new API to merge debug locations has been committed at r289661 (see
review D26256 for more details), we can use it to "improve" the code added by
revision r280995.

Instead of nulling the debugloc of a commoned instruction, we use the 'merged'
debug location. At the moment, this is just a no functional change since
function `DILocation::getMergedLocation()` is just a stub and would always
return a null location.

Differential Revision: https://reviews.llvm.org/D27804

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289862 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LiveRangeEdit] Change eliminateDeadDef assert to if condition.
Geoff Berry [Thu, 15 Dec 2016 19:55:19 +0000 (19:55 +0000)]
[LiveRangeEdit] Change eliminateDeadDef assert to if condition.

The assert could potentially fire (though no cases have been
encountered), so just check that the instruction we're handling
specially for rematerialization only has one def to begin with.

Reviewed by Wei Mi over email.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289861 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLibDriver: Allow resource files to be archive members.
Peter Collingbourne [Thu, 15 Dec 2016 19:37:46 +0000 (19:37 +0000)]
LibDriver: Allow resource files to be archive members.

It seems pointless to add a resource to an archive because it won't have
any symbols to link against (and link.exe doesn't have an equivalent of
--whole-archive), but lib.exe allows it for some reason.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289859 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRe-add the check for __has_attribute in StringLiteral.
Zachary Turner [Thu, 15 Dec 2016 19:33:31 +0000 (19:33 +0000)]
Re-add the check for __has_attribute in StringLiteral.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289858 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoBrainF example: fixing segfault caused by outdated code with missing MCJIT dependency
Boris Ulasevich [Thu, 15 Dec 2016 19:29:42 +0000 (19:29 +0000)]
BrainF example: fixing segfault caused by outdated code with missing MCJIT dependency
Differential Revision: https://reviews.llvm.org/D26280

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289857 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIgnore -Wgcc-compat diagnostic in StringLiteral.
Zachary Turner [Thu, 15 Dec 2016 19:22:58 +0000 (19:22 +0000)]
Ignore -Wgcc-compat diagnostic in StringLiteral.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289856 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] add folds for icmp (smin X, Y), X
Sanjay Patel [Thu, 15 Dec 2016 19:13:37 +0000 (19:13 +0000)]
[InstCombine] add folds for icmp (smin X, Y), X

Min/max canonicalization (r287585) exposes the fact that we're missing combines for min/max patterns.
This patch won't solve the example that was attached to that thread, so something else still needs fixing.

The line between InstCombine and InstSimplify gets blurry here because sometimes the icmp instruction that
we want to fold to already exists, but sometimes it's the swapped form of what we want.

Corresponding changes for smax/umin/umax to follow.

Differential Revision: https://reviews.llvm.org/D27531

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289855 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix some remaining documentation references to MSVC 2013
Reid Kleckner [Thu, 15 Dec 2016 19:08:02 +0000 (19:08 +0000)]
Fix some remaining documentation references to MSVC 2013

MSVC 2015 has been the minimum supported version of VS since October.

Differential Revision: https://reviews.llvm.org/D25710

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289854 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[StringRef] Add enable-if to StringLiteral.
Zachary Turner [Thu, 15 Dec 2016 19:02:43 +0000 (19:02 +0000)]
[StringRef] Add enable-if to StringLiteral.

to prevent StringLiteral from being created with a non-literal
char array, clang has a macro enable_if() that can be used
in such a way as to guarantee that the constructor is disabled
unless the length fo the string can be computed at compile time.

This only works on clang, but at least it should allow bots
to catch abuse of StringLiteral.

Differential Revision: https://reviews.llvm.org/D27780

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289853 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] doc update
Kostya Serebryany [Thu, 15 Dec 2016 18:47:22 +0000 (18:47 +0000)]
[libFuzzer] doc update

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289849 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GlobalISel] Drop workaround for Legalizer member/class sharing a name. NFC.
Ahmed Bougacha [Thu, 15 Dec 2016 18:45:30 +0000 (18:45 +0000)]
[GlobalISel] Drop workaround for Legalizer member/class sharing a name. NFC.

MachineLegalizer used to be the name of both the class and the member,
causing GCC errors. r276522 fixed that by renaming the member to just
'Legalizer'.  The 'class' workaround isn't necessary anymore; drop it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289848 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] use a single shufps for 256-bit vectors when it can save instructions
Sanjay Patel [Thu, 15 Dec 2016 18:43:46 +0000 (18:43 +0000)]
[x86] use a single shufps for 256-bit vectors when it can save instructions

This is the 256-bit counterpart to the 128-bit transform checked in here:
https://reviews.llvm.org/rL289837

This patch is based on the draft by @sroland (Roland Scheidegger) that is
attached to PR27885:
https://llvm.org/bugs/show_bug.cgi?id=27885

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289846 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Guard Misaligned 128-bit store penalty by subtarget feature
Matthew Simpson [Thu, 15 Dec 2016 18:36:59 +0000 (18:36 +0000)]
[AArch64] Guard Misaligned 128-bit store penalty by subtarget feature

This patch checks that the SlowMisaligned128Store subtarget feature is set
when penalizing such stores in getMemoryOpCost.

Differential Revision: https://reviews.llvm.org/D27677

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289845 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64][GlobalISel] Remove redundant RBI comments. NFC.
Ahmed Bougacha [Thu, 15 Dec 2016 18:22:15 +0000 (18:22 +0000)]
[AArch64][GlobalISel] Remove redundant RBI comments. NFC.

It's brittle, and Doxygen already picks the overriden method's comment
anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289844 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Ensure callees get hot threshold when first seen on cold path
Teresa Johnson [Thu, 15 Dec 2016 18:21:01 +0000 (18:21 +0000)]
[ThinLTO] Ensure callees get hot threshold when first seen on cold path

This is split out from D27696, since it turned out to be a bug fix and
not part of the NFC efficiency change.

Keep the same adjusted (possibly decayed) threshold in both the worklist
and the ImportList. Otherwise if we encountered it first along a cold
path, the callee would be added to the worklist with a lower decayed
threshold than when it is later encountered along a hot path. But the
logic uses the threshold recorded in the ImportList entry to check if
we should re-add it, and without this patch the threshold recorded there
is the same along both paths so we don't re-add it. Using the
same possibly decayed threshold in the ImportList ensures we re-add it
later with the higher non-decayed hot path threshold.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289843 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CMake] Minor change to symlink generation for LLDB
Chris Bieneman [Thu, 15 Dec 2016 18:17:07 +0000 (18:17 +0000)]
[CMake] Minor change to symlink generation for LLDB

If OUTPUT_DIR is not specified we can assume the symlink is linking to a file in the same directory, so we can use $<TARGET_FILE_NAME:${target}> to create a relative symlink.

In the case of LLDB, when we build a framework, we are creating symlinks in a different directory than the file we're pointing to, and we don't install those links. To make this work in the build directory we can use $<TARGET_FILE:${target}> instead, which uses the full path to the target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289840 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] use a single shufps when it can save instructions
Sanjay Patel [Thu, 15 Dec 2016 18:03:38 +0000 (18:03 +0000)]
[x86] use a single shufps when it can save instructions

This is a tiny patch with a big pile of test changes.
This partially fixes PR27885:
https://llvm.org/bugs/show_bug.cgi?id=27885

My motivating case looks like this:

  - vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,2]
  - vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
  - vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]

  + vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]

And this happens several times in the diffs. For chips with domain-crossing penalties,
the instruction count and size reduction should usually overcome any potential
domain-crossing penalty due to using an FP op in a sequence of int ops. For chips such
as recent Intel big cores and Atom, there is no domain-crossing penalty for shufps, so
using shufps is a pure win.

So the test case diffs all appear to be improvements except one test in
vector-shuffle-combining.ll where we miss an opportunity to use a shift to generate
zero elements and one test in combine-sra.ll where multiple uses prevent the expected
shuffle combining.

Differential Revision: https://reviews.llvm.org/D27692

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289837 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] Fix domains for scalar store instructions
Simon Pilgrim [Thu, 15 Dec 2016 17:09:24 +0000 (17:09 +0000)]
[X86][SSE] Fix domains for scalar store instructions

As discussed on D27692

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289834 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "[SimplifyCFG] In sinkLastInstruction correctly set debugloc of common inst"
Robert Lougher [Thu, 15 Dec 2016 16:59:13 +0000 (16:59 +0000)]
Revert "[SimplifyCFG] In sinkLastInstruction correctly set debugloc of common inst"

Reverting as it is causing buildbot failures (address sanitizer).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289833 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[lanai] Simplify small section check in LowerGlobalAddress and treat ldata sections...
Jacques Pienaar [Thu, 15 Dec 2016 16:56:16 +0000 (16:56 +0000)]
[lanai] Simplify small section check in LowerGlobalAddress and treat ldata sections specially.

Move the check for the code model into isGlobalInSmallSectionImpl and return false (not in small section) for variables placed in sections prefixed with .ldata (workaround for a tool limitation).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289832 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][AVX512] Moved instruction domain lookups to the right table. NFCI.
Simon Pilgrim [Thu, 15 Dec 2016 16:38:51 +0000 (16:38 +0000)]
[X86][AVX512] Moved instruction domain lookups to the right table. NFCI.

Avoid duplicating instructions in the int32/int64 domains.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289830 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SimplifyCFG] In sinkLastInstruction correctly set debugloc of "common" inst
Robert Lougher [Thu, 15 Dec 2016 16:17:53 +0000 (16:17 +0000)]
[SimplifyCFG] In sinkLastInstruction correctly set debugloc of "common" inst

Simplify CFG will try to sink the last instruction in a series of basic blocks,
creating a "common" instruction in the successor block (sinkLastInstruction).
When it does this, the debug location of the single instruction should be the
merged debug locations of the commoned instructions.

Differential Revision: https://reviews.llvm.org/D27590

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289828 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix ubsan failures in lane mask shifts
Krzysztof Parzyszek [Thu, 15 Dec 2016 16:08:49 +0000 (16:08 +0000)]
Fix ubsan failures in lane mask shifts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289826 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] Fix domains for VZEXT_LOAD type instructions
Simon Pilgrim [Thu, 15 Dec 2016 16:05:29 +0000 (16:05 +0000)]
[X86][SSE] Fix domains for VZEXT_LOAD type instructions

Add the missing domain equivalences for movss, movsd, movd and movq zero extending loading instructions.

Differential Revision: https://reviews.llvm.org/D27684

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289825 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix for regression after Global Load Scalarization patch
Alexander Timofeev [Thu, 15 Dec 2016 15:17:19 +0000 (15:17 +0000)]
Fix for regression after Global Load Scalarization patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289822 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoExtract LaneBitmask into a separate type
Krzysztof Parzyszek [Thu, 15 Dec 2016 14:36:06 +0000 (14:36 +0000)]
Extract LaneBitmask into a separate type

Specifically avoid implicit conversions from/to integral types to
avoid potential errors when changing the underlying type. For example,
a typical initialization of a "full" mask was "LaneMask = ~0u", which
would result in a value of 0x00000000FFFFFFFF if the type was extended
to uint64_t.

Differential Revision: https://reviews.llvm.org/D27454

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289820 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CostModel][X86] Updated reverse shuffle costs
Simon Pilgrim [Thu, 15 Dec 2016 14:24:07 +0000 (14:24 +0000)]
[CostModel][X86] Updated reverse shuffle costs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289819 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[TEST] Initial commit of tests for minmax horizontal reductions.
Alexey Bataev [Thu, 15 Dec 2016 13:21:29 +0000 (13:21 +0000)]
[TEST] Initial commit of tests for minmax horizontal reductions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289817 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "[TESTS] Initial commit of tests, by Andrew Tischenko"
Alexey Bataev [Thu, 15 Dec 2016 12:26:18 +0000 (12:26 +0000)]
Revert "[TESTS] Initial commit of tests, by Andrew Tischenko"

This reverts commit ee709f8988653a0334fbf100cdbbdd83a3933347.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289814 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] New opportunities for FoldAndOfICmp and FoldXorOfICmp
Ehsan Amiri [Thu, 15 Dec 2016 12:25:13 +0000 (12:25 +0000)]
[InstCombine] New opportunities for FoldAndOfICmp and FoldXorOfICmp

A number of new patterns for simplifying and/xor of icmp:

(icmp ne %x, 0) ^ (icmp ne %y, 0) => icmp ne %x, %y if the following is true:
1- (%x = and %a, %mask) and (%y = and %b, %mask)
2- %mask is a power of 2.

(icmp eq %x, 0) & (icmp ne %y, 0) => icmp ult %x, %y if the following is true:
1- (%x = and %a, %mask1) and (%y = and %b, %mask2)
2- Let %t be the smallest power of 2 where %mask1 & %t != 0. Then for any
   %s that is a power of 2 and %s & %mask2 != 0, we must have %s <= %t.
For example if %mask1 = 24 and %mask2 = 16, setting %s = 16 and %t = 8
violates condition (2) above. So this optimization cannot be applied.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289813 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CostModel] Fix long standing bug with reverse shuffle mask detection
Simon Pilgrim [Thu, 15 Dec 2016 12:12:45 +0000 (12:12 +0000)]
[CostModel] Fix long standing bug with reverse shuffle mask detection

Incorrect 'undef' mask index matching meant that broadcast shuffles could be detected as reverse shuffles

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289811 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[TESTS] Initial commit of tests, by Andrew Tischenko
Alexey Bataev [Thu, 15 Dec 2016 11:48:24 +0000 (11:48 +0000)]
[TESTS] Initial commit of tests, by Andrew Tischenko

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289807 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Power9] Allow AnyExt immediates for XXSPLTIB
Nemanja Ivanovic [Thu, 15 Dec 2016 11:16:20 +0000 (11:16 +0000)]
[Power9] Allow AnyExt immediates for XXSPLTIB

In some situations, the BUILD_VECTOR node that builds a v18i8 vector by
a splat of an i8 constant will end up with signed 8-bit values and other
situations, it'll end up with unsigned ones. Handle both situations.

Fixes PR31340.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289804 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVR] Support floats in the instrumention pass
Dylan McKay [Thu, 15 Dec 2016 11:02:41 +0000 (11:02 +0000)]
[AVR] Support floats in the instrumention pass

This also refactors some common code into the 'GetTypeName' method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289803 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CostModel][X86] Add tests for reverse shuffle costs
Simon Pilgrim [Thu, 15 Dec 2016 10:45:53 +0000 (10:45 +0000)]
[CostModel][X86] Add tests for reverse shuffle costs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289800 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd missing triple target for numeric section flag test
Prakhar Bahuguna [Thu, 15 Dec 2016 10:20:48 +0000 (10:20 +0000)]
Add missing triple target for numeric section flag test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289798 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSimplify format member detection in FormatVariadic
Pavel Labath [Thu, 15 Dec 2016 09:40:27 +0000 (09:40 +0000)]
Simplify format member detection in FormatVariadic

Summary:
This replaces the format member search, which was quite complicated, with a more
direct approach to detecting whether a class should be formatted using the
format-member method. Instead we use a special type llvm::format_adapter, which
every adapter must inherit from. Then the search can be simply implemented with
the is_base_of type trait.

Aside from the simplification, I like this way more because it makes it more
explicit that you are supposed to use this type only for adapter-like
formattings, and the other approach (format_provider overloads) should be used
as a default (a mistake I made when first trying to use this library).

The only slight change in behaviour here is that now choose the format-adapter
branch even if the format member invocation will fail to compile (e.g. because it is a
non-const member function and we are passing a const adapter), whereas
previously we would have gone on to search for format_providers for the type.
However, I think that is actually a good thing, as it probably means the
programmer did something wrong.

Reviewers: zturner, inglorion

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27679

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289795 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Thumb] Teach ISel how to lower compares of AND bitmasks efficiently
Sjoerd Meijer [Thu, 15 Dec 2016 09:38:59 +0000 (09:38 +0000)]
[Thumb] Teach ISel how to lower compares of AND bitmasks efficiently

This is essentially a recommit of r285893, but with a correctness fix. The
problem of the original commit was that this:

bic r5, r7, #31
cbz r5, .LBB2_10

got rewritten into:

lsrs  r5, r7, #5
beq .LBB2_10

The result in destination register r5 is not the same and this is incorrect
when r5 is not dead. So this fix includes checking the uses of the AND
destination register. And also, compared to the original commit, some regression
tests didn't need changing anymore because of this extra check.

For completeness, this was the original commit message:

For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more
efficient instruction selection if the bitmask is one consecutive sequence of
set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).

1) If the bitmask touches the LSB, then we can remove all the upper bits and
set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and
set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit
into the sign bit with one LSLS and change the condition query from NE/EQ to
MI/PL (we could also implement this by shifting into the carry bit and
branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower
zero bits of the mask.

1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two
16-bit instructions but can elide the CMP and doesn't require materializing a
complex immediate, so is also a win.

Differential Revision: https://reviews.llvm.org/D27761

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289794 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVR] Add argument indices to the instrumention hook functions
Dylan McKay [Thu, 15 Dec 2016 09:38:09 +0000 (09:38 +0000)]
[AVR] Add argument indices to the instrumention hook functions

This allows the instrumention hook functions to do better
pretty-printing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289793 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix for build warning in execute-only support
Prakhar Bahuguna [Thu, 15 Dec 2016 08:42:04 +0000 (08:42 +0000)]
Fix for build warning in execute-only support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289788 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAllow ELF section flags to be specified numerically
Prakhar Bahuguna [Thu, 15 Dec 2016 07:59:15 +0000 (07:59 +0000)]
Allow ELF section flags to be specified numerically

Summary:
GAS already allows flags for sections to be specified directly as a
numeric value. This functionality is particularly useful for setting
processor or application-specific values that may not be directly
supported or understood by LLVM. This patch allows LLVM to use numeric
section flag values verbatim if specified by the assembly file.

Reviewers: grosbach, rafael, t.p.northover, rengolin

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27451

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289785 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Implement execute-only support in CodeGen
Prakhar Bahuguna [Thu, 15 Dec 2016 07:59:08 +0000 (07:59 +0000)]
[ARM] Implement execute-only support in CodeGen

This implements execute-only support for ARM code generation, which
prevents the compiler from generating data accesses to code sections.
The following changes are involved:

* Add the CodeGen option "-arm-execute-only" to the ARM code generator.
* Add the clang flag "-mexecute-only" as well as the GCC-compatible
  alias "-mpure-code" to enable this option.
* When enabled, literal pools are replaced with MOVW/MOVT instructions,
  with VMOV used in addition for floating-point literals. As the MOVT
  instruction is required, execute-only support is only available in
  Thumb mode for targets supporting ARMv8-M baseline or Thumb2.
* Jump tables are placed in data sections when in execute-only mode.
* The execute-only text section is assigned section ID 0, and is
  marked as unreadable with the SHF_ARM_PURECODE flag with symbol 'y'.
  This also overrides selection of ELF sections for globals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289784 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd missing -mtriple to MIR test case
Sanjoy Das [Thu, 15 Dec 2016 07:13:50 +0000 (07:13 +0000)]
Add missing -mtriple to MIR test case

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289779 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAttempt to fix llvm-readobj crash on ppc64 due to r289674
Yaxun Liu [Thu, 15 Dec 2016 06:59:23 +0000 (06:59 +0000)]
Attempt to fix llvm-readobj crash on ppc64 due to r289674

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289777 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix go bindings after r289702 (hopefully, don't really know how to build
Daniel Jasper [Thu, 15 Dec 2016 06:54:29 +0000 (06:54 +0000)]
Fix go bindings after r289702 (hopefully, don't really know how to build
them, build.sh seems to be broken).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289775 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] enable the failure-resistant merge by default (with trace-pc-guard only)
Kostya Serebryany [Thu, 15 Dec 2016 06:21:21 +0000 (06:21 +0000)]
[libFuzzer] enable the failure-resistant merge by default (with trace-pc-guard only)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289772 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVR] Whitelist the avrlit config environment variables
Dylan McKay [Thu, 15 Dec 2016 06:04:53 +0000 (06:04 +0000)]
[AVR] Whitelist the avrlit config environment variables

This allows us to use `lit` to run on-target execution tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289769 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert part of r289765 that is not necessary
Hal Finkel [Thu, 15 Dec 2016 05:50:45 +0000 (05:50 +0000)]
Revert part of r289765 that is not necessary

CS.doesNotAccessMemory(ArgNo) and CS.onlyReadsMemory(ArgNo) calls
dataOperandHasImpliedAttr, so revert this part of r289765 because
it should not be necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289768 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTrying to fix NDEBUG build after r289764
Hal Finkel [Thu, 15 Dec 2016 05:33:19 +0000 (05:33 +0000)]
Trying to fix NDEBUG build after r289764

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289766 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix argument attribute queries with bundle operands
Hal Finkel [Thu, 15 Dec 2016 05:09:15 +0000 (05:09 +0000)]
Fix argument attribute queries with bundle operands

When iterating over data operands in AA, don't make argument-attribute-specific
queries on bundle operands. Trying to fix self hosting...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289765 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MachineBlockPlacement] Don't make blocks "uneditable"
Sanjoy Das [Thu, 15 Dec 2016 05:08:57 +0000 (05:08 +0000)]
[MachineBlockPlacement] Don't make blocks "uneditable"

Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true.  The timeline is as
follows:

 1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
    `TailDup.shouldTailDuplicate` on its argument, which in turn calls
    `analyzeBranch` with `AllowModify` set to true.

 2. This `analyzeBranch` call edits the terminator sequence of the block
    based on the physical layout of the machine function, turning an
    unanalyzable non-fallthrough block to a unanalyzable fallthrough
    block.  Normally MBP bails out of rearranging such blocks, but this
    block was unanalyzable non-fallthrough (and thus rearrangeable) the
    first time MBP looked at it, and so it goes ahead and decides where
    it should be placed in the function.

 3. When placing this block MBP fails to analyze and thus update the
    block in keeping with the new physical layout.

Concretely, before (1) we have something like:

```
LBL0:
  < unknown terminator op that may branch to LBL1 >
  jmp LBL1

LBL1:
  ... A

LBL2:
  ... B
```

In (2), analyze branch simplifies this to

```
LBL0:
  < unknown terminator op that may branch to LBL2 >
  ;; jmp LBL1 <- redundant jump removed

LBL1:
  ... A

LBL2:
  ... B
```

In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.

```
LBL0:
  < unknown terminator op that may branch to LBL2 >
  ;; jmp LBL1 <- redundant jump

LBL2:
  ... B

LBL1:
  ... A
```

and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.

There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator.  That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.

I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.

Reviewers: chandlerc, gberry, MatzeB, iteratee

Subscribers: mcrosier, llvm-commits

Differential Revision: https://reviews.llvm.org/D27783

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512][InstCombine] Add masked scalar FMA intrinsics to SimplifyDemandedVectorElts.
Craig Topper [Thu, 15 Dec 2016 03:49:45 +0000 (03:49 +0000)]
[AVX-512][InstCombine] Add masked scalar FMA intrinsics to SimplifyDemandedVectorElts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289759 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix iterator-invalidation issue
Hal Finkel [Thu, 15 Dec 2016 03:30:40 +0000 (03:30 +0000)]
Fix iterator-invalidation issue

Inserting a new key into a DenseMap potentially invalidates iterators into that
map. Trying to fix an issue from r289755 triggering this assertion:

  Assertion `isHandleInSync() && "invalid iterator access!"' failed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289757 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove the AssumptionCache
Hal Finkel [Thu, 15 Dec 2016 03:02:15 +0000 (03:02 +0000)]
Remove the AssumptionCache

After r289755, the AssumptionCache is no longer needed. Variables affected by
assumptions are now found by using the new operand-bundle-based scheme. This
new scheme is more computationally efficient, and also we need much less
code...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289756 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMake processing @llvm.assume more efficient by using operand bundles
Hal Finkel [Thu, 15 Dec 2016 02:53:42 +0000 (02:53 +0000)]
Make processing @llvm.assume more efficient by using operand bundles

There was an efficiency problem with how we processed @llvm.assume in
ValueTracking (and other places). The AssumptionCache tracked all of the
assumptions in a given function. In order to find assumptions relevant to
computing known bits, etc. we searched every assumption in the function. For
ValueTracking, that means that we did O(#assumes * #values) work in InstCombine
and other passes (with a constant factor that can be quite large because we'd
repeat this search at every level of recursion of the analysis).

Several of us discussed this situation at the last developers' meeting, and
this implements the discussed solution: Make the values that an assume might
affect operands of the assume itself. To avoid exposing this detail to
frontends and passes that need not worry about it, I've used the new
operand-bundle feature to add these extra call "operands" in a way that does
not affect the intrinsic's signature. I think this solution is relatively
clean. InstCombine adds these extra operands based on what ValueTracking, LVI,
etc. will need and then those passes need only search the users of the values
under consideration. This should fix the computational-complexity problem.

At this point, no passes depend on the AssumptionCache, and so I'll remove
that as a follow-up change.

Differential Revision: https://reviews.llvm.org/D27259

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289755 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd testcases for some shuffle bugs.
Eli Friedman [Thu, 15 Dec 2016 01:47:15 +0000 (01:47 +0000)]
Add testcases for some shuffle bugs.

See https://llvm.org/bugs/show_bug.cgi?id=31301 and
https://llvm.org/bugs/show_bug.cgi?id=31364 .

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289751 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix test/tools/lto/hide-linkonce-odr.ll after r289719
Nico Weber [Thu, 15 Dec 2016 01:31:38 +0000 (01:31 +0000)]
Fix test/tools/lto/hide-linkonce-odr.ll after r289719

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289750 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Remove dead #defines from NVPTXUtilities.h.
Justin Lebar [Thu, 15 Dec 2016 00:45:06 +0000 (00:45 +0000)]
[NVPTX] Remove dead #defines from NVPTXUtilities.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289747 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse PIC relocation model as default for PowerPC64 ELF.
Joerg Sonnenberger [Thu, 15 Dec 2016 00:01:53 +0000 (00:01 +0000)]
Use PIC relocation model as default for PowerPC64 ELF.

Most of the PowerPC64 code generation for the ELF ABI is already PIC.
There are four main exceptions:
(1) Constant pointer arrays etc. should in writeable sections.
(2) The TOC restoration NOP after a call is needed for all global
symbols. While GNU ld has a workaround for questionable GCC self-calls,
we trigger the checks for calls from COMDAT sections as they cross input
sections and are therefore not considered self-calls. The current
decision is questionable and suboptimal, but outside the scope of the
change.
(3) TLS access can not use the initial-exec model.
(4) Jump tables should use relative addresses. Note that the current
encoding doesn't work for the large code model, but it is more compact
than the default for any non-trivial jump table. Improving this is again
beyond the scope of this change.

At least (1) and (3) are assumptions made in target-independent code and
introducing additional hooks is a bit messy. Testing with clang shows
that a -fPIC binary is 600KB smaller than the corresponding -fno-pic
build. Separate testing from improved jump table encodings would explain
only about 100KB or so. The rest is expected to be a result of more
aggressive immediate forming for -fno-pic, where the -fPIC binary just
uses TOC entries.

This change brings the LLVM output in line with the GCC output, other
PPC64 compilers like XLC on AIX are known to produce PIC by default
as well. The relocation model can still be provided explicitly, i.e.
when using MCJIT.

One test case for case (1) is included, other test cases with relocation
mode sensitive behavior are wired to static for now. They will be
reviewed and adjusted separately.

Differential Revision: https://reviews.llvm.org/D26566

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289743 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] Fix runtime-metadata.ll test so it doesn't leave an object file in the sourc...
Justin Lebar [Wed, 14 Dec 2016 23:24:43 +0000 (23:24 +0000)]
[AMDGPU] Fix runtime-metadata.ll test so it doesn't leave an object file in the source tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289742 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Remove dead code.
Justin Lebar [Wed, 14 Dec 2016 23:20:40 +0000 (23:20 +0000)]
[NVPTX] Remove dead code.

I've chosen to remove NVPTXInstrInfo::CanTailMerge but not
NVPTXInstrInfo::isLoadInstr and isStoreInstr (which are also dead)
because while the latter two are reasonably useful utilities, the former
cannot be used safely: It relies on successful address space inference
to identify writes to shared memory, but addrspace inference is a
best-effort thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289740 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAG] allow more select folding for targets that have 'and not' (PR31175)
Sanjay Patel [Wed, 14 Dec 2016 22:59:14 +0000 (22:59 +0000)]
[DAG] allow more select folding for targets that have 'and not' (PR31175)

The original motivation for this patch comes from wanting to canonicalize
more IR to selects and also canonicalizing min/max.

If we're going to do that, we need more backend fixups to undo select codegen
when simpler ops will do. I chose AArch64 for the tests because that shows the
difference in the simplest way. This should fix:
https://llvm.org/bugs/show_bug.cgi?id=31175

Differential Revision: https://reviews.llvm.org/D27489

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289738 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[gold] Add datalayout to two tests where it was missing.
Davide Italiano [Wed, 14 Dec 2016 22:53:43 +0000 (22:53 +0000)]
[gold] Add datalayout to two tests where it was missing.

Reported by: thakis via chromium bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289737 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Wed, 14 Dec 2016 22:50:46 +0000 (22:50 +0000)]
[Hexagon] 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@289736 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd the ability to get attribute values as Optional<T>
Greg Clayton [Wed, 14 Dec 2016 22:38:08 +0000 (22:38 +0000)]
Add the ability to get attribute values as Optional<T>

When getting attributes it is sometimes nicer to use Optional<T> some of the time instead of magic values. I tried to cut over to only using the Optional values but it made many of the call sites very messy, so it makes sense the leave in the calls that can return a default value. Otherwise code that looks like this:

uint64_t CallColumn = Die.getAttributeValueAsAddress(DW_AT_call_line, 0);

Has to be turned into:

uint64_t CallColumn = 0;
if (auto CallColumnValue = Die.getAttributeValueAsAddress(DW_AT_call_line))
    CallColumn = *CallColumnValue;

The first snippet of code looks much better. But in cases where you want an offset that may or may not be there, the following code looks better:

if (auto StmtOffset = Die.getAttributeValueAsSectionOffset(DW_AT_stmt_list)) {
  // Use StmtOffset
}

Differential Revision: https://reviews.llvm.org/D27772

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289731 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWhitespace cleanup in test/CodeGen/NVPTX/annotations.ll.
Justin Lebar [Wed, 14 Dec 2016 22:32:55 +0000 (22:32 +0000)]
Whitespace cleanup in test/CodeGen/NVPTX/annotations.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289730 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Support .maxnreg annotation.
Justin Lebar [Wed, 14 Dec 2016 22:32:50 +0000 (22:32 +0000)]
[NVPTX] Support .maxnreg annotation.

Reviewers: tra

Subscribers: llvm-commits, jholewinski

Differential Revision: https://reviews.llvm.org/D27638

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289729 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Remove string constants from NVPTXBaseInfo.h.
Justin Lebar [Wed, 14 Dec 2016 22:32:44 +0000 (22:32 +0000)]
[NVPTX] Remove string constants from NVPTXBaseInfo.h.

Summary:
Previously they were defined as a 2D char array in a header file.  This
is kind of overkill -- we can let the linker lay out these strings
however it pleases.  While we're at it, we might as well just inline
these constants where they're used, as each of them is used only once.

Also move NVPTXUtilities.{h,cpp} into namespace llvm.

Reviewers: tra

Subscribers: jholewinski, mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D27636

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289728 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLibDriver: Reject inputs that are not COFF objects or bitcode files.
Peter Collingbourne [Wed, 14 Dec 2016 22:19:22 +0000 (22:19 +0000)]
LibDriver: Reject inputs that are not COFF objects or bitcode files.

Fixes PR31372.

Differential Revision: https://reviews.llvm.org/D27776

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289726 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoOnly sets profile summary when it was not preset.
Dehao Chen [Wed, 14 Dec 2016 22:06:49 +0000 (22:06 +0000)]
Only sets profile summary when it was not preset.

Summary: SampleProfileLoader pass may be invoked twice by LTO. The 2nd pass should not append more summary info as it is already preset by the 1st pass.

Reviewers: eraman, davidxl

Subscribers: mehdi_amini, llvm-commits

Differential Revision: https://reviews.llvm.org/D27733

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289725 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix the bug in r289714 (NFC).
Dehao Chen [Wed, 14 Dec 2016 22:03:08 +0000 (22:03 +0000)]
Fix the bug in r289714 (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289724 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert revision 289721.
Jan Sjodin [Wed, 14 Dec 2016 21:58:42 +0000 (21:58 +0000)]
Revert revision 289721.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289723 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDummy commit.
Jan Sjodin [Wed, 14 Dec 2016 21:57:18 +0000 (21:57 +0000)]
Dummy commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289721 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LTO] Add the missing datalayout in a test.
Davide Italiano [Wed, 14 Dec 2016 21:57:14 +0000 (21:57 +0000)]
[LTO] Add the missing datalayout in a test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289720 91177308-0d34-0410-b5e6-96231b3b80d8