platform/upstream/llvm.git
7 years ago[X86][XOP] Merge rotation opcodes with AVX512 equivalents. NFCI.
Simon Pilgrim [Tue, 26 Sep 2017 14:12:50 +0000 (14:12 +0000)]
[X86][XOP] Merge rotation opcodes with AVX512 equivalents. NFCI.

The XOP rotations act as ROTL with +ve values and ROTR with -ve values, which means that we can treat them all as ROTL with unsigned modulo. We already check that we're only trying to lower as ROTL for XOP rotations.

Differential Revision: https://reviews.llvm.org/D37949

llvm-svn: 314207

7 years ago[DSE] Merge stores when the later store only writes to memory locations the early...
Sanjay Patel [Tue, 26 Sep 2017 13:54:28 +0000 (13:54 +0000)]
[DSE] Merge stores when the later store only writes to memory locations the early store also wrote to (2nd try)

This is a 2nd attempt at:
https://reviews.llvm.org/rL310055
...which was reverted at rL310123 because of PR34074:
https://bugs.llvm.org/show_bug.cgi?id=34074

In this version, we break out of the inner loop after we successfully merge and kill a pair of stores. In the
earlier rev, we were continuing instead, which meant we could process the invalid info from a now dead store.

Original commit message (authored by Filipe Cabecinhas):

This fixes PR31777.

If both stores' values are ConstantInt, we merge the two stores
(shifting the smaller store appropriately) and replace the earlier (and
larger) store with an updated constant.

In the future we should also support vectors of integers. And maybe
float/double if we can.

Differential Revision: https://reviews.llvm.org/D30703

llvm-svn: 314206

7 years ago[OPENMP] Generate implicit map|firstprivate clauses for target-based
Alexey Bataev [Tue, 26 Sep 2017 13:47:31 +0000 (13:47 +0000)]
[OPENMP] Generate implicit map|firstprivate clauses for target-based
directives.

If the variable is used in the target-based region but is not found in
any private|mapping clause, then generate implicit firstprivate|map
clauses for these implicitly mapped variables.

llvm-svn: 314205

7 years ago[ELF] - Remove wrong comment from testcase.
George Rimar [Tue, 26 Sep 2017 13:35:00 +0000 (13:35 +0000)]
[ELF] - Remove wrong comment from testcase.

It was introduced by me in D37059.
Comment was saying that Weak binding is incorrect
for 'foo' symbol and that should be true for symbol in final output.
But at that place LTO temporarily file was checked,
where Weak binding for 'foo' is fine as LTO changes binding for
'LinkerRedefined' symbols internally to prevent IPO.

Binding for 'foo' in final output is correctly set to Global
and that tested just few lines below in the same testcase.

llvm-svn: 314204

7 years ago[x86] fix pr29061
Coby Tayree [Tue, 26 Sep 2017 13:28:05 +0000 (13:28 +0000)]
[x86] fix pr29061

https://bugs.llvm.org//show_bug.cgi?id=29061
Don't try referencing REX-needed regs when not on 64bit mode
Aligns to GCC

Differetial Revision: https://reviews.llvm.org/D37801

llvm-svn: 314203

7 years agoTidyup P->getComplexPatternInfo call by moving it inside if( != NULL) test. NFCI.
Simon Pilgrim [Tue, 26 Sep 2017 12:59:01 +0000 (12:59 +0000)]
Tidyup P->getComplexPatternInfo call by moving it inside if( != NULL) test. NFCI.

llvm-svn: 314202

7 years agoDon't move llvm.localescape outside the entry block in the GCOV profiling pass
Sylvestre Ledru [Tue, 26 Sep 2017 11:56:43 +0000 (11:56 +0000)]
Don't move llvm.localescape outside the entry block in the GCOV profiling pass

Summary:
This fixes https://bugs.llvm.org/show_bug.cgi?id=34714.

Patch by Marco Castelluccio

Reviewers: rnk

Reviewed By: rnk

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D38224

llvm-svn: 314201

7 years ago[mips][asan] Fix preinstalled_signal.cc test for mips
Simon Dardis [Tue, 26 Sep 2017 11:49:04 +0000 (11:49 +0000)]
[mips][asan] Fix preinstalled_signal.cc test for mips

Linux for mips has a non-standard layout for the kernel sigaction struct.
Adjust the layout by the minimally amount to get the test to pass, as we
don't require the usage of the restorer function.

llvm-svn: 314200

7 years agoRevert "[X86] Make all the NOREX CodeGenOnly instructions into postRA pseudos like...
Benjamin Kramer [Tue, 26 Sep 2017 10:25:27 +0000 (10:25 +0000)]
Revert "[X86] Make all the NOREX CodeGenOnly instructions into postRA pseudos like the NOREX version of TEST."

Makes llc crash. This reverts commit r314151.

llvm-svn: 314199

7 years ago[dsymutil] Better support for symbol aliases
Jonas Devlieghere [Tue, 26 Sep 2017 08:17:28 +0000 (08:17 +0000)]
[dsymutil] Better support for symbol aliases

This patch adds logic to follow a symbol's aliases when the symbol name
cannot be found in the current object file. It checks the main binary
for the symbol's address and queries the current object for its aliases
(symbols with the same address) before printing out a warning.

Differential revision: https://reviews.llvm.org/D38230

llvm-svn: 314198

7 years agoSkip building unused parts when targeting SJLJ
Martin Storsjo [Tue, 26 Sep 2017 08:07:26 +0000 (08:07 +0000)]
Skip building unused parts when targeting SJLJ

When SJLJ exceptions are used, those functions aren't used.

This fixes build failures on ARM with SJLJ enabled (e.g. on armv7/iOS)
when built using the CMake project files.

Differential Revision: https://reviews.llvm.org/D38249

llvm-svn: 314197

7 years agoCorrect data types in the _Unwind_FunctionContext struct
Martin Storsjo [Tue, 26 Sep 2017 08:07:17 +0000 (08:07 +0000)]
Correct data types in the _Unwind_FunctionContext struct

This makes it match the definition used within llvm and in libgcc,
we previously got the wrong layout in 64 bit environments.

Differential Revision: https://reviews.llvm.org/D38247

llvm-svn: 314196

7 years ago[X86] Finishing broadcastf32x2 and broadcasti32x2 intrinsics lowering to IR. llvm...
Uriel Korach [Tue, 26 Sep 2017 07:39:39 +0000 (07:39 +0000)]
[X86] Finishing broadcastf32x2 and broadcasti32x2 intrinsics lowering to IR. llvm side.

Removing X86 broadcast(f/i)32x2 intrinsics from llvm.
Adding autoUpgrade support.
Moving matching tests from avx512dq-intrinsics.ll to avx512dq-intrinsics-upgrade.ll and from avx512dqvl-intrinsics.ll to avx512dqvl-intrinsics-upgrade.ll.

Differential Revision: https://reviews.llvm.org/D38220

llvm-svn: 314195

7 years ago[XRay] Avoid actual linking when testing the driver
Dean Michael Berris [Tue, 26 Sep 2017 04:29:25 +0000 (04:29 +0000)]
[XRay] Avoid actual linking when testing the driver

Use -### in the command to see just look for the output of -v.

Follow-up to D38226.

llvm-svn: 314194

7 years ago[XRay] Run command once without piping to FileCheck
Dean Michael Berris [Tue, 26 Sep 2017 04:17:37 +0000 (04:17 +0000)]
[XRay] Run command once without piping to FileCheck

This allows us to debug the failures that come up from the build bots.

Follow-up to D38226.

llvm-svn: 314193

7 years agoSpeed up SymbolTable::insert().
Rui Ueyama [Tue, 26 Sep 2017 04:17:13 +0000 (04:17 +0000)]
Speed up SymbolTable::insert().

SymbolTable::insert() is a hot path function. When linking a clang debug
build, the function is called 3.7 million times. The total amount of "Name"
string contents is 300 MiB. That means this `Name.find("@@")` scans almost
300 MiB of data. That's far from negligible.

StringRef::find(StringRef) uses a sophisticated algorithm, but the
function is slow for a short needle. This patch replaces it with
StringRef::find(char).

This patch alone speeds up a clang debug build link time by 0.5 seconds
from 8.2s to 7.7s. That's 6% speed up. It seems too good for this tiny
change, but looks like it's real.

llvm-svn: 314192

7 years ago[XRay] Only run shared tests when 'enable_shared' is true
Dean Michael Berris [Tue, 26 Sep 2017 04:07:45 +0000 (04:07 +0000)]
[XRay] Only run shared tests when 'enable_shared' is true

Follow-up to D38226.

llvm-svn: 314191

7 years ago[XRay] Remove -fPIC from shared build test.
Dean Michael Berris [Tue, 26 Sep 2017 04:00:41 +0000 (04:00 +0000)]
[XRay] Remove -fPIC from shared build test.

Follow-up to D38226.

llvm-svn: 314190

7 years ago[XRay] Stop running tests for 'amd64', and remove -fPIE from tests.
Dean Michael Berris [Tue, 26 Sep 2017 03:45:37 +0000 (03:45 +0000)]
[XRay] Stop running tests for 'amd64', and remove -fPIE from tests.

Follow-up to D38226.

llvm-svn: 314189

7 years ago[XRay][Driver] Do not link in XRay runtime in shared libs
Dean Michael Berris [Tue, 26 Sep 2017 03:18:11 +0000 (03:18 +0000)]
[XRay][Driver] Do not link in XRay runtime in shared libs

Summary:
This change ensures that we don't link in the XRay runtime when building
shared libraries with clang. This doesn't prevent us from building
shared libraris tht have XRay instrumentation sleds, but it does prevent
us from linking in the static XRay runtime into a shared library.

The XRay runtime currently doesn't support dynamic registration of
instrumentation sleds in shared objects, which we'll start enabling in
the future. That work has to happen in the back-end and in the runtime.

Reviewers: rnk, pelikan

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D38226

llvm-svn: 314188

7 years agoCodeGenModule: Adapt to LLVM TargetLibraryInfo changes
Matthias Braun [Tue, 26 Sep 2017 02:37:23 +0000 (02:37 +0000)]
CodeGenModule: Adapt to LLVM TargetLibraryInfo changes

Adapt to LLVM TargetLibraryInfo changes in r314185.

See also https://reviews.llvm.org/D38106 and https://reviews.llvm.org/D37891

llvm-svn: 314187

7 years agoCMake: Add option to set LLVM_ENABLE_DUMP
Matthias Braun [Tue, 26 Sep 2017 02:36:58 +0000 (02:36 +0000)]
CMake: Add option to set LLVM_ENABLE_DUMP

Differential Revision: https://reviews.llvm.org/D38267

llvm-svn: 314186

7 years agoTargetLibraryInfo: Stop guessing wchar_t size
Matthias Braun [Tue, 26 Sep 2017 02:36:57 +0000 (02:36 +0000)]
TargetLibraryInfo: Stop guessing wchar_t size

Usually the frontend communicates the size of wchar_t via metadata and
we can optimize wcslen (and possibly other calls in the future). In
cases without the wchar_size metadata we would previously try to guess
the correct size based on the target triple; however this is fragile to
keep up to date and may miss users manually changing the size via flags.
Better be safe and stop guessing and optimizing if the frontend didn't
communicate the size.

Differential Revision: https://reviews.llvm.org/D38106

llvm-svn: 314185

7 years ago[AVR] Fix the build after setting alignment to 1 in r314179
Dylan McKay [Tue, 26 Sep 2017 02:07:54 +0000 (02:07 +0000)]
[AVR] Fix the build after setting alignment to 1 in r314179

Changing all types to be byte-aligned broke a small number of tests.

llvm-svn: 314183

7 years ago[AVR] Prefer BasicBlock::getIterator over Function::begin()
Dylan McKay [Tue, 26 Sep 2017 01:37:53 +0000 (01:37 +0000)]
[AVR] Prefer BasicBlock::getIterator over Function::begin()

Thanks to Eli Friedman for the suggestion.

llvm-svn: 314182

7 years agoSplit MergeSyntheticSection into Merge{Tail,NoTail}Section.
Rui Ueyama [Tue, 26 Sep 2017 00:54:24 +0000 (00:54 +0000)]
Split MergeSyntheticSection into Merge{Tail,NoTail}Section.

This patch alone is neutral in terms of code readability, but this
change makes a following patch easier to read.

llvm-svn: 314181

7 years ago[AVR] When lowering shifts into loops, put newly generated MBBs in the same
Dylan McKay [Tue, 26 Sep 2017 00:51:03 +0000 (00:51 +0000)]
[AVR] When lowering shifts into loops, put newly generated MBBs in the same
spot as the original MBB

Discovered in avr-rust/rust#62
https://github.com/avr-rust/rust/issues/62

Patch by Gergo Erdi.

llvm-svn: 314180

7 years ago[AVR] Use 1-byte alignment for all data types
Dylan McKay [Tue, 26 Sep 2017 00:45:27 +0000 (00:45 +0000)]
[AVR] Use 1-byte alignment for all data types

This was an oversight in the original backend data layout.

The AVR architecture does not have the concept of unaligned loads - all
loads/stores from all addresses are aligned to one byte.

Discovered in avr-rust issue #64
https://github.com/avr-rust/rust/issues/64

Patch By Gergo Erdi.

llvm-svn: 314179

7 years agoRevert "[XRay][Driver] Do not link in XRay runtime in shared libs"
Dean Michael Berris [Tue, 26 Sep 2017 00:41:08 +0000 (00:41 +0000)]
Revert "[XRay][Driver] Do not link in XRay runtime in shared libs"

Reverts r314177.

llvm-svn: 314178

7 years ago[XRay][Driver] Do not link in XRay runtime in shared libs
Dean Michael Berris [Mon, 25 Sep 2017 23:40:33 +0000 (23:40 +0000)]
[XRay][Driver] Do not link in XRay runtime in shared libs

Summary:
This change ensures that we don't link in the XRay runtime when building
shared libraries with clang. This doesn't prevent us from building
shared libraris tht have XRay instrumentation sleds, but it does prevent
us from linking in the static XRay runtime into a shared library.

The XRay runtime currently doesn't support dynamic registration of
instrumentation sleds in shared objects, which we'll start enabling in
the future. That work has to happen in the back-end and in the runtime.

Reviewers: rnk, pelikan

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D38226

llvm-svn: 314177

7 years ago[docs] llvm-cov: Make docs for boolean options more consistent
Vedant Kumar [Mon, 25 Sep 2017 23:10:04 +0000 (23:10 +0000)]
[docs] llvm-cov: Make docs for boolean options more consistent

llvm-svn: 314176

7 years ago[llvm-cov] Warn if -show-functions is used without query files
Vedant Kumar [Mon, 25 Sep 2017 23:10:03 +0000 (23:10 +0000)]
[llvm-cov] Warn if -show-functions is used without query files

llvm-cov's report mode does not print any output when -show-functions is
specified and no source files are specified. This can be surprising, so
the tool should at least print out an error message when this happens.

rdar://problem/34636859

llvm-svn: 314175

7 years agoModernize comments
Adrian Prantl [Mon, 25 Sep 2017 22:51:26 +0000 (22:51 +0000)]
Modernize comments

llvm-svn: 314174

7 years agoModernize comments
Adrian Prantl [Mon, 25 Sep 2017 22:51:15 +0000 (22:51 +0000)]
Modernize comments

llvm-svn: 314173

7 years agoclang-format/java: Always put space after `assert` keyword.
Nico Weber [Mon, 25 Sep 2017 22:42:49 +0000 (22:42 +0000)]
clang-format/java: Always put space after `assert` keyword.

Previously, it was missing if the expression after the assert started with a (.

llvm-svn: 314172

7 years agoAllow specifying sanitizers in blacklists
Vlad Tsyrklevich [Mon, 25 Sep 2017 22:11:12 +0000 (22:11 +0000)]
Allow specifying sanitizers in blacklists

Summary:
This is the follow-up patch to D37924.

This change refactors clang to use the the newly added section headers
in SpecialCaseList to specify which sanitizers blacklists entries
should apply to, like so:

  [cfi-vcall]
  fun:*bad_vcall*
  [cfi-derived-cast|cfi-unrelated-cast]
  fun:*bad_cast*

The SanitizerSpecialCaseList class has been added to allow querying by
SanitizerMask, and SanitizerBlacklist and its downstream users have been
updated to provide that information. Old blacklists not using sections
will continue to function identically since the blacklist entries will
be placed into a '[*]' section by default matching against all
sanitizers.

Reviewers: pcc, kcc, eugenis, vsk

Reviewed By: eugenis

Subscribers: dberris, cfe-commits, mgorny

Differential Revision: https://reviews.llvm.org/D37925

llvm-svn: 314171

7 years agoAdd section headers to SpecialCaseLists
Vlad Tsyrklevich [Mon, 25 Sep 2017 22:11:11 +0000 (22:11 +0000)]
Add section headers to SpecialCaseLists

Summary:
Sanitizer blacklist entries currently apply to all sanitizers--there
is no way to specify that an entry should only apply to a specific
sanitizer. This is important for Control Flow Integrity since there are
several different CFI modes that can be enabled at once. For maximum
security, CFI blacklist entries should be scoped to only the specific
CFI mode(s) that entry applies to.

Adding section headers to SpecialCaseLists allows users to specify more
information about list entries, like sanitizer names or other metadata,
like so:

  [section1]
  fun:*fun1*
  [section2|section3]
  fun:*fun23*

The section headers are regular expressions. For backwards compatbility,
blacklist entries entered before a section header are put into the '[*]'
section so that blacklists without sections retain the same behavior.

SpecialCaseList has been modified to also accept a section name when
matching against the blacklist. It has also been modified so the
follow-up change to clang can define a derived class that allows
matching sections by SectionMask instead of by string.

Reviewers: pcc, kcc, eugenis, vsk

Reviewed By: eugenis, vsk

Subscribers: vitalybuka, llvm-commits

Differential Revision: https://reviews.llvm.org/D37924

llvm-svn: 314170

7 years agoRevert r312724 ("[ARM] Remove redundant vcvt patterns.").
Eli Friedman [Mon, 25 Sep 2017 22:07:33 +0000 (22:07 +0000)]
Revert r312724 ("[ARM] Remove redundant vcvt patterns.").

It leads to some improvements, but also a regression for the simple
case, so it's not clearly a good idea.

test/CodeGen/ARM/vcvt.ll now has test coverage to show the difference.

Ultimately, the right solution is probably to custom-lower fp-to-int
conversions, to something like ARMISD::VCVT_F32_S32 plus a bitcast.
It's hard to do the right thing when the implicit bitcast isn't visible
to DAG transforms.

llvm-svn: 314169

7 years ago[GlobalISel] Update the documentation and comment for G_[UN]MERGE_VALUES
Quentin Colombet [Mon, 25 Sep 2017 22:03:06 +0000 (22:03 +0000)]
[GlobalISel] Update the documentation and comment for G_[UN]MERGE_VALUES

In r296921, we added the G_[UN]MERGE_VALUES node, but did not update the
documentation. Fixing that.

NFC.

llvm-svn: 314168

7 years ago[GlobalISel] Update the documentation for G_SEQUENCE
Quentin Colombet [Mon, 25 Sep 2017 22:03:05 +0000 (22:03 +0000)]
[GlobalISel] Update the documentation for G_SEQUENCE

This instruction has been removed in r306120.

NFC.

llvm-svn: 314167

7 years ago[GlobalISel] Update the documentation and comments for G_EXTRACT
Quentin Colombet [Mon, 25 Sep 2017 22:03:01 +0000 (22:03 +0000)]
[GlobalISel] Update the documentation and comments for G_EXTRACT

In r297100, G_EXTRACT changed from a multiple results instruction to a
single result one. Update the documentation accordingly.

NFC.

llvm-svn: 314166

7 years agoX86: remove R12 from CSR on Windows x64 SwiftCC
Saleem Abdulrasool [Mon, 25 Sep 2017 22:00:17 +0000 (22:00 +0000)]
X86: remove R12 from CSR on Windows x64 SwiftCC

R12 is used for the SwiftError parameter.  It is no longer a CSR as it
is used for transfer the SwiftError, and the caller must preserve it if
they need to.

llvm-svn: 314165

7 years ago[OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain.
Gheorghe-Teodor Bercea [Mon, 25 Sep 2017 21:56:32 +0000 (21:56 +0000)]
[OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain.

Summary: Enable the -nocudalib flag for the OpenMP device offloading toolchain as well. Currently it can only be used for the CUDA toolchain.

Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, hfinkel, tra

Reviewed By: tra

Subscribers: hfinkel, cfe-commits

Differential Revision: https://reviews.llvm.org/D37913

llvm-svn: 314164

7 years ago[ARM] Fix tests for vcvt+store to return void.
Eli Friedman [Mon, 25 Sep 2017 21:55:27 +0000 (21:55 +0000)]
[ARM] Fix tests for vcvt+store to return void.

This is what I meant to do in r314161; I didn't realize I'd messed up
because the generated assembly is currently identical.

llvm-svn: 314163

7 years agoAdd missing include to sanitizer procmaps
Francis Ricci [Mon, 25 Sep 2017 21:51:04 +0000 (21:51 +0000)]
Add missing include to sanitizer procmaps

llvm-svn: 314162

7 years ago[ARM] Add tests for vcvt followed by store.
Eli Friedman [Mon, 25 Sep 2017 21:37:52 +0000 (21:37 +0000)]
[ARM] Add tests for vcvt followed by store.

llvm-svn: 314161

7 years ago[ARM] Regenerate vcvt test checks.
Eli Friedman [Mon, 25 Sep 2017 21:34:29 +0000 (21:34 +0000)]
[ARM] Regenerate vcvt test checks.

llvm-svn: 314160

7 years ago[x86] make assertions less strict in avx512f test file
Sanjay Patel [Mon, 25 Sep 2017 21:31:08 +0000 (21:31 +0000)]
[x86] make assertions less strict in avx512f test file

Missed a line in r314158.

llvm-svn: 314159

7 years ago[x86] make assertions less strict in avx512f test file
Sanjay Patel [Mon, 25 Sep 2017 21:27:37 +0000 (21:27 +0000)]
[x86] make assertions less strict in avx512f test file

I'm not sure why yet, but there may be differences depending on the host?

llvm-svn: 314158

7 years agoFix style in sanitizer_procmaps.h
Francis Ricci [Mon, 25 Sep 2017 21:26:34 +0000 (21:26 +0000)]
Fix style in sanitizer_procmaps.h

llvm-svn: 314157

7 years ago[OpenMP] Bugfix: output file name drops the absolute path where full path is needed.
Gheorghe-Teodor Bercea [Mon, 25 Sep 2017 21:25:38 +0000 (21:25 +0000)]
[OpenMP] Bugfix: output file name drops the absolute path where full path is needed.

Summary: When composing the output file name, the path to the file is being dropped. The full path is required.

Reviewers: Hahnfeld, ABataev, caomhin, carlo.bertolli, hfinkel, tra

Reviewed By: tra

Subscribers: hfinkel, tra, cfe-commits

Differential Revision: https://reviews.llvm.org/D37912

llvm-svn: 314156

7 years agoRemove the details of the libstdc++ implementation that were
Jason Molenda [Mon, 25 Sep 2017 21:23:37 +0000 (21:23 +0000)]
Remove the details of the libstdc++ implementation that were
in TestDataFormatterSkipSummary.py - I'm building this test
with the default c++ library.

Skip TestMTCSimple.py when running for i386.

llvm-svn: 314155

7 years agoRevert commit with wrong message.
Gheorghe-Teodor Bercea [Mon, 25 Sep 2017 21:22:49 +0000 (21:22 +0000)]
Revert commit with wrong message.

llvm-svn: 314154

7 years ago[InstCombine] Move an optimization from foldICmpAndConstConst to foldICmpUsingKnownBits
Craig Topper [Mon, 25 Sep 2017 21:15:00 +0000 (21:15 +0000)]
[InstCombine] Move an optimization from foldICmpAndConstConst to foldICmpUsingKnownBits

All this optimization cares about is knowing how many low bits of LHS is known to be zero and whether that means that the result is 0 or greater than the RHS constant. It doesn't matter where the zeros in the low bits came from. So we don't need to specifically look for an AND. Instead we can use known bits.

Differential Revision: https://reviews.llvm.org/D38195

llvm-svn: 314153

7 years ago[X86] Don't select anyext GR32->GR64 to SUBREG_TO_REG. Use INSERT_SUBREG instead.
Craig Topper [Mon, 25 Sep 2017 21:14:59 +0000 (21:14 +0000)]
[X86] Don't select anyext GR32->GR64 to SUBREG_TO_REG. Use INSERT_SUBREG instead.

As far as I know SUBREG_TO_REG is stating that the upper bits are 0. But if we are just converting the GR32 with no checks, then we have no reason to say the upper bits are 0.

I don't really know how to test this today since I can't find anything that looks that closely at SUBREG_TO_REG. The test changes here seems to be some perturbance of register allocation.

Differential Revision: https://reviews.llvm.org/D38001

llvm-svn: 314152

7 years ago[X86] Make all the NOREX CodeGenOnly instructions into postRA pseudos like the NOREX...
Craig Topper [Mon, 25 Sep 2017 21:14:55 +0000 (21:14 +0000)]
[X86] Make all the NOREX CodeGenOnly instructions into postRA pseudos like the NOREX version of TEST.

llvm-svn: 314151

7 years ago[OpenMP] Don't throw cudalib not found error if only front-end is required.
Gheorghe-Teodor Bercea [Mon, 25 Sep 2017 21:07:16 +0000 (21:07 +0000)]
[OpenMP] Don't throw cudalib not found error if only front-end is required.

Summary: If we only use the compiler front-end, do not throw an error about the cuda device library not being found. This allows the front-end to be run on systems where no Cuda installation is found.

Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, tra

Reviewed By: tra

Subscribers: hfinkel, tra, cfe-commits

Differential Revision: https://reviews.llvm.org/D37914

llvm-svn: 314150

7 years ago[sanitizer_common] Don't provide sanitizer_procmaps API functions where not defined
Francis Ricci [Mon, 25 Sep 2017 20:48:51 +0000 (20:48 +0000)]
[sanitizer_common] Don't provide sanitizer_procmaps API functions where not defined

Summary:
Platforms that don't implement procmaps (primarily fuchsia and windows) still expose
the procmaps API when including sanitizer_procmaps.h, despite not implementing the functions
provided by that header. Ensure that the API is only exposed on platforms that implement it.

Reviewers: vitalybuka, alekseyshl, kubamracek

Subscribers: llvm-commits, krytarowski

Differential Revision: https://reviews.llvm.org/D38187

llvm-svn: 314149

7 years ago[llvm-objcopy] Refactor code to include initialize method
Jake Ehrlich [Mon, 25 Sep 2017 20:37:28 +0000 (20:37 +0000)]
[llvm-objcopy] Refactor code to include initialize method

This change refactors some of the code to allow for some code
deduplication in later diffs as well as just to make adding a new
section type more self contained to the class itself. The idea for this
was first mentioned by James in D 37915 and will be used in that change
as recommended.

This change follows changes for dynamic sections but precedes support
for dynamic relocations.

Differential Revision: https://reviews.llvm.org/D38008

llvm-svn: 314148

7 years ago[InstCombine] remove extract-of-select vector transform (2nd try)
Sanjay Patel [Mon, 25 Sep 2017 20:30:53 +0000 (20:30 +0000)]
[InstCombine] remove extract-of-select vector transform (2nd try)

The 1st attempt at this:
https://reviews.llvm.org/rL314117
was reverted at:
https://reviews.llvm.org/rL314118

because of bot fails for clang tests that were checking optimized IR. That should be fixed with:
https://reviews.llvm.org/rL314144
...so try again.

Original commit message:

The transform to convert an extract-of-a-select-of-vectors was added at:
https://reviews.llvm.org/rL194013

And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>

Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.

The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.

The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.

Differential Revision: https://reviews.llvm.org/D38006

llvm-svn: 314147

7 years ago[ScopInfo] Allow invariant loads in branch conditions
Tobias Grosser [Mon, 25 Sep 2017 20:27:15 +0000 (20:27 +0000)]
[ScopInfo] Allow invariant loads in branch conditions

In case the value used in a branch condition is a load instruction, assume this
load to be invariant.

llvm-svn: 314146

7 years agoReinstall the patch "Use EmitPointerWithAlignment to get alignment information of...
Wei Mi [Mon, 25 Sep 2017 19:57:59 +0000 (19:57 +0000)]
Reinstall the patch "Use EmitPointerWithAlignment to get alignment information of the pointer used in atomic expr" after fixing PR31620.

This is to fix PR34347. EmitAtomicExpr now only uses alignment information from
Type, instead of Decl, so when the declaration of an atomic variable is marked
to have the alignment equal as its size, EmitAtomicExpr doesn't know about it and
will generate libcall instead of atomic op. The patch uses EmitPointerWithAlignment
to get the precise alignment information.

Differential Revision: https://reviews.llvm.org/D37310

llvm-svn: 314145

7 years ago[x86] remove RUNs that were checking fully optimized IR
Sanjay Patel [Mon, 25 Sep 2017 19:56:57 +0000 (19:56 +0000)]
[x86] remove RUNs that were checking fully optimized IR

Clang regression tests that depend on the optimizer can break
when there are changes to LLVM...as in:
https://reviews.llvm.org/rL314117

llvm-svn: 314144

7 years ago[Hexagon] Avoid unused variable warnings in Release builds.
Benjamin Kramer [Mon, 25 Sep 2017 19:42:20 +0000 (19:42 +0000)]
[Hexagon] Avoid unused variable warnings in Release builds.

No functionality change intended.

llvm-svn: 314143

7 years agoRevert "[NVPTX] added match.{any,all}.sync instructions, intrinsics & builtins."...
Justin Lebar [Mon, 25 Sep 2017 19:41:56 +0000 (19:41 +0000)]
Revert "[NVPTX] added match.{any,all}.sync instructions, intrinsics & builtins.", rL314135.

Causing assertion failures on macos:

> Assertion failed: (Num < NumOperands && "Invalid child # of SDNode!"),
> function getOperand, file
> /Users/buildslave/jenkins/workspace/clang-stage1-cmake-RA-incremental/llvm/include/llvm/CodeGen/SelectionDAGNodes.h,
> line 835.

http://green.lab.llvm.org/green/job/clang-stage1-cmake-RA-incremental/42739/testReport/LLVM/CodeGen_NVPTX/surf_read_cuda_ll/

llvm-svn: 314142

7 years ago[analyzer] Fix crash on modeling of pointer arithmetic
Alexander Shaposhnikov [Mon, 25 Sep 2017 19:32:33 +0000 (19:32 +0000)]
[analyzer] Fix crash on modeling of pointer arithmetic

This patch fixes analyzer's crash on the newly added test case
(see also https://bugs.llvm.org/show_bug.cgi?id=34374).
Pointers subtraction appears to be modeled incorrectly
in the following example:
  char* p;
  auto n = p - reinterpret_cast<char*>((unsigned long)1);
In this case the analyzer (built without this patch)
tries to create a symbolic value for the difference
treating reinterpret_cast<char*>((unsigned long)1)
as an integer, that is not correct.

Differential revision: https://reviews.llvm.org/D38214

Test plan: make check-all

llvm-svn: 314141

7 years ago[X86] [ASM INTEL SYNTAX] fix for incorrect assembler code generation when x86-asm...
Konstantin Belochapka [Mon, 25 Sep 2017 19:26:48 +0000 (19:26 +0000)]
[X86] [ASM INTEL SYNTAX] fix for incorrect assembler code generation when x86-asm-syntax=intel (PR34617).
Fix for incorrect code generation when x86-asm-syntax=intel.
Differential Revision: https://reviews.llvm.org/D37945

llvm-svn: 314140

7 years ago[SelectionDAG] Teach simplifyDemandedBits to handle shifts by constant splat vectors
Craig Topper [Mon, 25 Sep 2017 19:26:08 +0000 (19:26 +0000)]
[SelectionDAG] Teach simplifyDemandedBits to handle shifts by constant splat vectors

This teach simplifyDemandedBits to handle constant splat vector shifts.

This required changing some uses of getZExtValue to getLimitedValue since we can't rely on legalization using getShiftAmountTy for the shift amount.

I believe there may have been a bug in the ((X << C1) >>u ShAmt) handling where we didn't check if the inner shift was too large. I've fixed that here.

I had to add new patterns to ARM because the zext/sext the patterns were trying to look for got turned into an any_extend with this patch. Happy to split that out too, but not sure how to test without this change.

Differential Revision: https://reviews.llvm.org/D37665

llvm-svn: 314139

7 years ago[MinGW] Don't link -lmsvcrt if a different msvcrt version is to be linked
Martin Storsjo [Mon, 25 Sep 2017 19:24:45 +0000 (19:24 +0000)]
[MinGW] Don't link -lmsvcrt if a different msvcrt version is to be linked

Differential Revision: https://reviews.llvm.org/D37530

llvm-svn: 314138

7 years ago[SLP] Add a test for PR32086, NFC.
Alexey Bataev [Mon, 25 Sep 2017 19:12:59 +0000 (19:12 +0000)]
[SLP] Add a test for PR32086, NFC.

llvm-svn: 314137

7 years ago[Hexagon] Better determination of register classes in bit tracker
Krzysztof Parzyszek [Mon, 25 Sep 2017 19:12:55 +0000 (19:12 +0000)]
[Hexagon] Better determination of register classes in bit tracker

Add two callbacks to MachineEvaluator, so that specific implementations
can specify more details about register classes:
- composeWithSubRegIndex(RC,Idx), to provide the register class for a
  register from RC used in conjunction with a subregister index Idx.
- getPhysRegBitWidth(Reg), to provide the size in bits of the given
  physical register.

llvm-svn: 314136

7 years ago[NVPTX] added match.{any,all}.sync instructions, intrinsics & builtins.
Artem Belevich [Mon, 25 Sep 2017 18:53:57 +0000 (18:53 +0000)]
[NVPTX] added match.{any,all}.sync instructions, intrinsics & builtins.

Differential Revision: https://reviews.llvm.org/D38191

llvm-svn: 314135

7 years ago[Hexagon] Make getHexagonSubRegIndex take reference instead of pointer
Krzysztof Parzyszek [Mon, 25 Sep 2017 18:49:42 +0000 (18:49 +0000)]
[Hexagon] Make getHexagonSubRegIndex take reference instead of pointer

llvm-svn: 314134

7 years ago[AVX-512] Replace large number of explicit patterns that check for insert_subvector...
Craig Topper [Mon, 25 Sep 2017 18:43:13 +0000 (18:43 +0000)]
[AVX-512] Replace large number of explicit patterns that check for insert_subvector with zero after masked compares with fewer patterns with predicate

This replaces the large number of patterns that handle every possible case of zeroing after a masked compare with a few simpler patterns that use a predicate to check for a masked compare producer.

This is similar to what we do for detecting free GR32->GR64 zero extends and free xmm->ymm/zmm zero extends.

This shrinks the isel table from ~590k to ~531k. This is a roughly 10% reduction in size.

Differential Revision: https://reviews.llvm.org/D38217

llvm-svn: 314133

7 years agoInitial patchset to get the testsuite running against armv7 and arm64 iOS devices.
Jason Molenda [Mon, 25 Sep 2017 18:19:39 +0000 (18:19 +0000)]
Initial patchset to get the testsuite running against armv7 and arm64 iOS devices.
Normal customer devices won't be able to run these tests, we're hoping to get
a public facing bot set up at some point.  Both devices pass the testsuite without
any errors or failures.

I have seen some instability with the armv7 test runs, I may submit additional patches
to address this.  arm64 looks good.

I'll be watching the bots for the rest of today; if any problems are introduced by
this patch I'll revert it - if anyone sees a problem with their bot that I don't
see, please do the same.  I know it's a rather large patch.

One change I had to make specifically for iOS devices was that debugserver can't
create files.  There were several tests that launch the inferior process redirecting
its output to a file, then they retrieve the file.  They were not trying to test
file redirection in these tests, so I rewrote those to write their output to a file
directly.

llvm-svn: 314132

7 years ago[SimplifyIndvar] Minor change to refine r314125, NFC
Hongbin Zheng [Mon, 25 Sep 2017 18:10:36 +0000 (18:10 +0000)]
[SimplifyIndvar] Minor change to refine r314125, NFC

llvm-svn: 314130

7 years ago[CUDA] Fix names of __nvvm_vote* intrinsics.
Artem Belevich [Mon, 25 Sep 2017 17:55:26 +0000 (17:55 +0000)]
[CUDA] Fix names of __nvvm_vote* intrinsics.

Also fixed a syntax error in activemask().

Differential Revision: https://reviews.llvm.org/D38188

llvm-svn: 314129

7 years agoARM: One more fix for swifterror CSR set
Arnold Schwaighofer [Mon, 25 Sep 2017 17:51:33 +0000 (17:51 +0000)]
ARM: One more fix for swifterror CSR set

We use a differently ordered CSR set if the frame pointer is pushed. Add a
matching ..._SwiftError version.

llvm-svn: 314128

7 years agoUse socketpair on all Unix platforms
Eugene Zemtsov [Mon, 25 Sep 2017 17:41:16 +0000 (17:41 +0000)]
Use socketpair on all Unix platforms

Using TCP sockets is insecure against local attackers, and possibly
against remote attackers too (some vulnerabilities may allow tricking a
browser to make a request to localhost). Use socketpair (which is immune
to such attacks) on all Unix platforms.

Patch by Demi Marie Obenour < demiobenour@gmail.com >

Differential Revision: https://reviews.llvm.org/D33213

llvm-svn: 314127

7 years agoUse UINT64_MAX instead of implicitly-type-converted -1.
Rui Ueyama [Mon, 25 Sep 2017 17:40:21 +0000 (17:40 +0000)]
Use UINT64_MAX instead of implicitly-type-converted -1.

llvm-svn: 314126

7 years ago[SimplifyIndvar] Replace the srem used by IV if we can prove both of its operands...
Hongbin Zheng [Mon, 25 Sep 2017 17:39:40 +0000 (17:39 +0000)]
[SimplifyIndvar] Replace the srem used by IV if we can prove both of its operands are non-negative

Since now SCEV can handle 'urem', an 'urem' is a better canonical form than an 'srem' because it has well-defined behavior

This is a follow up of D34598

Differential Revision: https://reviews.llvm.org/D38072

llvm-svn: 314125

7 years ago[Sema] Null check in BuildDeclarationNameExpr
Yi Kong [Mon, 25 Sep 2017 17:36:54 +0000 (17:36 +0000)]
[Sema] Null check in BuildDeclarationNameExpr

Qualtype may point to null if we cannot infer its type yet.

Fixes PR33843

Differential Revision: https://reviews.llvm.org/D38158

llvm-svn: 314124

7 years ago[ARM] Fix -Wdangling-else warning.
Benjamin Kramer [Mon, 25 Sep 2017 17:35:38 +0000 (17:35 +0000)]
[ARM] Fix -Wdangling-else warning.

A ternary is clearer here. No functionality change.

llvm-svn: 314123

7 years agoRevert "Initial patchset to get the testsuite running against armv7 and arm64 iOS...
Chris Bieneman [Mon, 25 Sep 2017 17:31:40 +0000 (17:31 +0000)]
Revert "Initial patchset to get the testsuite running against armv7 and arm64 iOS devices. Normal customer devices won't be able to run these devices, we're hoping to get a public facing bot set up at some point.  Both devices pass the testsuite without any errors or failures."

This patch has been causing LLDB test failures on ObjC tests. A test log
may still be available here:

http://lab.llvm.org:8080/green/view/LLDB/job/lldb/1650/

This reverts commit r314038.

llvm-svn: 314122

7 years agoARM: Use the proper swifterror CSR list on platforms other than darwin
Arnold Schwaighofer [Mon, 25 Sep 2017 17:19:50 +0000 (17:19 +0000)]
ARM: Use the proper swifterror CSR list on platforms other than darwin

Noticed by inspection

llvm-svn: 314121

7 years agoUse a temporary varaible to improve readability. NFC.
Rui Ueyama [Mon, 25 Sep 2017 17:19:17 +0000 (17:19 +0000)]
Use a temporary varaible to improve readability. NFC.

llvm-svn: 314120

7 years ago[clangd] Fix missing "message" key when responding with unsupported method
Benjamin Kramer [Mon, 25 Sep 2017 17:16:47 +0000 (17:16 +0000)]
[clangd] Fix missing "message" key when responding with unsupported method

The language server protocol dictates that a ResponseError should have a
[message string][1] describing the error. This adds a simple message to the
error and a simple test.

[1]: https://github.com/Microsoft/language-server-protocol/blob/master/protocol.md#response-message

Patch by Raoul Wols!

Differential Revision: https://reviews.llvm.org/D38225

llvm-svn: 314119

7 years agorevert r314117 because there are bogus clang tests that depend on the optimizer
Sanjay Patel [Mon, 25 Sep 2017 17:00:04 +0000 (17:00 +0000)]
revert r314117 because there are bogus clang tests that depend on the optimizer

llvm-svn: 314118

7 years ago[InstCombine] remove extract-of-select vector transform
Sanjay Patel [Mon, 25 Sep 2017 16:41:34 +0000 (16:41 +0000)]
[InstCombine] remove extract-of-select vector transform

The transform to convert an extract-of-a-select-of-vectors was added at:
rL194013

And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>

Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.

The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.

The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.

Differential Revision: https://reviews.llvm.org/D38006

llvm-svn: 314117

7 years ago[ScopInfo] Allow uniform branch conditions
Tobias Grosser [Mon, 25 Sep 2017 16:37:15 +0000 (16:37 +0000)]
[ScopInfo] Allow uniform branch conditions

If all but one branch come from an error condition and the incoming value from
this branch is a constant, we can model this branch.

llvm-svn: 314116

7 years agoRemove trailing whitespaces.
Michael Liao [Mon, 25 Sep 2017 16:21:21 +0000 (16:21 +0000)]
Remove trailing whitespaces.

llvm-svn: 314115

7 years ago[DebugInfo] Sort the SDDbgValue list before assuming it is in IR order
Reid Kleckner [Mon, 25 Sep 2017 16:14:53 +0000 (16:14 +0000)]
[DebugInfo] Sort the SDDbgValue list before assuming it is in IR order

Summary:
This code iterates the 'Orders' vector in parallel with the DbgValue
list, emitting all DBG_VALUEs that occurred between the last IR order
insertion point and the next insertion point. This assumes the
SDDbgValue list is sorted in IR order, which it usually is. However, it
is not sorted when a node with a debug value is replaced with another
one. When this happens, TransferDbgValues is called, and the new value
is added to the end of the list.

The problem can be solved by stably sorting the list by IR order.

Reviewers: aprantl, Ka-Ka

Reviewed By: aprantl

Subscribers: MatzeB, hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D38197

llvm-svn: 314114

7 years agoUse {} instead of make_pair and an iterator for the insertion point, NFC
Reid Kleckner [Mon, 25 Sep 2017 16:14:39 +0000 (16:14 +0000)]
Use {} instead of make_pair and an iterator for the insertion point, NFC

llvm-svn: 314113

7 years agoRework atomic ops to use clang builtins rather than llvm asm
Jan Vesely [Mon, 25 Sep 2017 16:07:34 +0000 (16:07 +0000)]
Rework atomic ops to use clang builtins rather than llvm asm

reviewer: Aaron Watry

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 314112

7 years agoprepare_builtins: Fix compile breakage with older LLVM
Jan Vesely [Mon, 25 Sep 2017 16:04:37 +0000 (16:04 +0000)]
prepare_builtins: Fix compile breakage with older LLVM

Fixes r314050

reviewer: Tom Stellard

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 314111

7 years ago[scudo] Scudo thread specific data refactor, part 2
Kostya Kortchinsky [Mon, 25 Sep 2017 15:12:08 +0000 (15:12 +0000)]
[scudo] Scudo thread specific data refactor, part 2

Summary:
Following D38139, we now consolidate the TSD definition, merging the shared
TSD definition with the exclusive TSD definition. We introduce a boolean set
at initializaton denoting the need for the TSD to be unlocked or not. This
adds some unused members to the exclusive TSD, but increases consistency and
reduces the definitions fragmentation.

We remove the fallback mechanism from `scudo_allocator.cpp` and add a fallback
TSD in the non-shared version. Since the shared version doesn't require one,
this makes overall more sense.

There are a couple of additional cosmetic changes: removing the header guards
from the remaining `.inc` files, added error string to a `CHECK`.

Question to reviewers: I thought about friending `getTSDAndLock` in `ScudoTSD`
so that the `FallbackTSD` could `Mutex.Lock()` directly instead of `lock()`
which involved zeroing out the `Precedence`, which is unused otherwise. Is it
worth doing?

Reviewers: alekseyshl, dvyukov, kcc

Reviewed By: dvyukov

Subscribers: srhines, llvm-commits

Differential Revision: https://reviews.llvm.org/D38183

llvm-svn: 314110

7 years ago[X86][LLVM]Expanding Supports lowerInterleavedStore() in X86InterleavedAccess (VF8...
Michael Zuckerman [Mon, 25 Sep 2017 14:50:38 +0000 (14:50 +0000)]
[X86][LLVM]Expanding Supports lowerInterleavedStore() in X86InterleavedAccess (VF8 stride 4):

This patch expands the support of lowerInterleavedStore to 8x8i stride 4.

LLVM creates suboptimal shuffle code-gen for AVX2.
In overall, this patch is a specific fix for the pattern (Strid=4 VF=8) and we plan to include more patterns in the future.

The patch goal is to optimize the following sequence:
At the end of the computation, we have xmm2, xmm0, xmm12 and xmm3 holding
each 8 chars:

c0, c1, , c7
m0, m1, , m7
y0, y1, , y7
k0, k1, ., k7

And these need to be transposed/interleaved and stored like so:

c0 m0 y0 k0 c1 m1 y1 k1 c2 m2 y2 k2 c3 m3 y3 k3 ....

Reviewers
DavidKreitzer
Farhana
zvi
igorb
guyblank
RKSimon
Ayal

Differential Revision: https://reviews.llvm.org/D36058

Change-Id: I3cc5c2ca5d6318901c192a4428493b99ef424c32
llvm-svn: 314109

7 years ago[ELF][fix] Using statically linked lld leads to segmentation fault on linking ELF
Evgeny Mankov [Mon, 25 Sep 2017 14:42:15 +0000 (14:42 +0000)]
[ELF][fix] Using statically linked lld leads to segmentation fault on linking ELF

[Synopsys]
Using function elf::link(...) leads to segmentation fault on its second call. First call finishes correctly.
[Solution]
Clear the rest of globals.

Reviewed by: George Rimar and Rui Ueyama

Differential Revision: http://reviews.llvm.org/D38131

llvm-svn: 314108

7 years agoFix clangd when built with LLVM_LINK_LLVM_DYLIB=ON
Sylvestre Ledru [Mon, 25 Sep 2017 14:08:35 +0000 (14:08 +0000)]
Fix clangd when built with LLVM_LINK_LLVM_DYLIB=ON

Reviewers: malaperle, malaperle-ericsson, bkramer

Reviewed By: bkramer

Subscribers: bkramer, mgorny, ilya-biryukov, cfe-commits

Differential Revision: https://reviews.llvm.org/D38228

llvm-svn: 314107

7 years ago[PowerPC] Eliminate compares - add i64 sext/zext handling for SETLT/SETGT
Nemanja Ivanovic [Mon, 25 Sep 2017 14:05:46 +0000 (14:05 +0000)]
[PowerPC] Eliminate compares - add i64 sext/zext handling for SETLT/SETGT

As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential review.

llvm-svn: 314106