platform/upstream/llvm.git
6 years agoSimplify getOffset for synthetic sections.
Rafael Espindola [Thu, 19 Apr 2018 16:54:30 +0000 (16:54 +0000)]
Simplify getOffset for synthetic sections.

We had a single symbol using -1 with a synthetic section. It is
simpler to just update its value.

This is not a big will by itself, but will allow having a simple
getOffset for InputSeciton.

llvm-svn: 330340

6 years agoRename MergeInputSection::getOffset.
Rafael Espindola [Thu, 19 Apr 2018 16:05:07 +0000 (16:05 +0000)]
Rename MergeInputSection::getOffset.

Unlike the getOffset in the base class, this one computes the offset
in the parent synthetic section, not the final output section.

llvm-svn: 330339

6 years ago[CXX] Templates specialization visibility can be wrong
Steven Wu [Thu, 19 Apr 2018 15:46:43 +0000 (15:46 +0000)]
[CXX] Templates specialization visibility can be wrong

Summary:
Under some conditions, LinkageComputer can get the visibility for
ClassTemplateSpecializationDecl wrong because it failed to find the Decl
that has the explicit visibility.

This fixes:
llvm.org/bugs/pr36810
rdar://problem/38080953

Reviewers: rsmith, arphaman, doug.gregor

Reviewed By: doug.gregor

Subscribers: doug.gregor, cfe-commits

Differential Revision: https://reviews.llvm.org/D44670

llvm-svn: 330338

6 years ago[AMDGPU] Do not only rely on BB number when finding bottom loop
Mark Searles [Thu, 19 Apr 2018 15:42:30 +0000 (15:42 +0000)]
[AMDGPU] Do not only rely on BB number when finding bottom loop

We should also check that the "bottom" basic block of a loopis a successor of the "header" basic block, otherwise we don't propagate the information correctly when the CFG is complex. This fixes an important rendering problem with Wolfsentein 2, because of one vector-memory wait was missing.

Differential Revision: https://reviews.llvm.org/D43831

llvm-svn: 330337

6 years ago[NEON] Define vfma_n_f32() and vfmaq_n_f32() intrinsics in AArch32 mode
Ivan A. Kosarev [Thu, 19 Apr 2018 15:27:28 +0000 (15:27 +0000)]
[NEON] Define vfma_n_f32() and vfmaq_n_f32() intrinsics in AArch32 mode

Differential Revision: https://reviews.llvm.org/D45670

llvm-svn: 330336

6 years ago[llvm-mca][X86] Add mmx instruction to btver2 resource tests
Simon Pilgrim [Thu, 19 Apr 2018 15:09:46 +0000 (15:09 +0000)]
[llvm-mca][X86] Add mmx instruction to btver2 resource tests

Useful to see scheduler class deltas against xmm equivalents

llvm-svn: 330335

6 years ago[NewGVN] Add ops as dependency if we cannot find a leader for ValueOp.
Florian Hahn [Thu, 19 Apr 2018 15:05:47 +0000 (15:05 +0000)]
[NewGVN] Add ops as dependency if we cannot find a leader for ValueOp.

If those operands change, we might find a leader for ValueOp, which
could enable new phi-of-op creation.

This fixes a case where we missed creating a phi-of-ops node. With D43865
and this patch, bootstrapping clang/llvm works with -enable-newgvn, whereas
without it, the "value changed after iteration" assertion is triggered.

Reviewers: dberlin, davide

Reviewed By: dberlin

Differential Revision: https://reviews.llvm.org/D42180

llvm-svn: 330334

6 years ago[Hexagon] Generate code for vector bswap intrinsics
Krzysztof Parzyszek [Thu, 19 Apr 2018 14:46:44 +0000 (14:46 +0000)]
[Hexagon] Generate code for vector bswap intrinsics

llvm-svn: 330333

6 years ago[X86][BtVer2] Remove SSE4A EXTRQ/EXTRQI InstRW overrides.
Simon Pilgrim [Thu, 19 Apr 2018 14:38:36 +0000 (14:38 +0000)]
[X86][BtVer2] Remove SSE4A EXTRQ/EXTRQI InstRW overrides.

These are already handled identically by WriteALU.

llvm-svn: 330332

6 years agoFix __attribute__((force_align_arg_pointer)) misalignment bug
Erich Keane [Thu, 19 Apr 2018 14:27:05 +0000 (14:27 +0000)]
Fix __attribute__((force_align_arg_pointer)) misalignment bug

The force_align_arg_pointer attribute was using a hardcoded 16-byte
alignment value which in combination with -mstack-alignment=32 (or
larger) would produce a misaligned stack which could result in crashes
when accessing stack buffers using aligned AVX load/store instructions.

Fix the issue by using the "stackrealign" function attribute instead
of using a hardcoded 16-byte alignment.

Patch By: Gramner

Differential Revision: https://reviews.llvm.org/D45812

llvm-svn: 330331

6 years ago[Hexagon] Add/fix patterns for 32/64-bit vector compares and logical ops
Krzysztof Parzyszek [Thu, 19 Apr 2018 14:24:31 +0000 (14:24 +0000)]
[Hexagon] Add/fix patterns for 32/64-bit vector compares and logical ops

llvm-svn: 330330

6 years ago[Unittests] Fix plugins test
Mikhail Maltsev [Thu, 19 Apr 2018 14:02:46 +0000 (14:02 +0000)]
[Unittests] Fix plugins test

Summary:
Currently the PluginsTests.LoadPlugin unit test is failing in
LLVM configurations that have LLVM_EXPORT_SYMBOLS_FOR_PLUGINS enabled
because the EnableABIBreakingChecks symbol is missing.

This patch fixes the issue by linking some additional libraries to the
test plugin if LLVM_EXPORT_SYMBOLS_FOR_PLUGINS is enabled.

Reviewers: philip.pfaffe

Reviewed By: philip.pfaffe

Subscribers: mgorny, llvm-commits, rogfer01

Differential Revision: https://reviews.llvm.org/D45811

llvm-svn: 330329

6 years ago[Fuzzer] Make InterruptHandler non-blocking for Fuchsia
Petr Hosek [Thu, 19 Apr 2018 14:01:46 +0000 (14:01 +0000)]
[Fuzzer] Make InterruptHandler non-blocking for Fuchsia

The initial naive approach to simulate SIGINT on Fuchsia was to getchar
and look for ETX. This caused the InterruptHandler thread to lock stdin,
preventing musl's exit() from being able to close the stdio descriptors
and complete. This change uses select() instead.

Patch By: aarongreen

Differential Revision: https://reviews.llvm.org/D45636

llvm-svn: 330328

6 years ago[clang-tidy] Fix unused-variable warning.
Haojian Wu [Thu, 19 Apr 2018 13:34:03 +0000 (13:34 +0000)]
[clang-tidy] Fix unused-variable warning.

llvm-svn: 330327

6 years ago[mips] Correct the definitions of the unaligned word memory operation instructions
Simon Dardis [Thu, 19 Apr 2018 13:33:51 +0000 (13:33 +0000)]
[mips] Correct the definitions of the unaligned word memory operation instructions

These instructions lacked the correct predicates, were not marked
as loads and stores and lacked the proper instruction mapping information.

In the case of microMIPS sw(l|r)e (EVA) these instructions were using the load
EVA description.

Reviewers: abeserminji, smaksimovic, atanasyan

Differential Revision: https://reviews.llvm.org/D45626

llvm-svn: 330326

6 years ago[NFC][InstCombine] A few more tests for masked merge add/xor -> or with constant...
Roman Lebedev [Thu, 19 Apr 2018 13:02:17 +0000 (13:02 +0000)]
[NFC][InstCombine] A few more tests for masked merge add/xor -> or with constant mask

llvm-svn: 330325

6 years ago[clang-format] Don't remove empty lines before namespace endings
Krasimir Georgiev [Thu, 19 Apr 2018 13:02:15 +0000 (13:02 +0000)]
[clang-format] Don't remove empty lines before namespace endings

Summary: This implements an alternative to r327861, namely preserving empty lines before namespace endings.

Reviewers: djasper

Reviewed By: djasper

Subscribers: klimek, cfe-commits

Differential Revision: https://reviews.llvm.org/D45373

llvm-svn: 330324

6 years agoLowering x86 adds/addus/subs/subus intrinsics (clang)
Alexander Ivchenko [Thu, 19 Apr 2018 12:15:11 +0000 (12:15 +0000)]
Lowering x86 adds/addus/subs/subus intrinsics (clang)

This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations.

Patch by tkrupa

Differential Revision: https://reviews.llvm.org/D44786

llvm-svn: 330323

6 years agoLowering x86 adds/addus/subs/subus intrinsics (llvm part)
Alexander Ivchenko [Thu, 19 Apr 2018 12:13:30 +0000 (12:13 +0000)]
Lowering x86 adds/addus/subs/subus intrinsics (llvm part)

This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations. The patch also includes folding
of previously missing saturation patterns so that IR emits the same
machine instructions as the intrinsics.

Patch by tkrupa

Differential Revision: https://reviews.llvm.org/D44785

llvm-svn: 330322

6 years agoRemove file accidentally added in r330320.
Florian Hahn [Thu, 19 Apr 2018 12:09:05 +0000 (12:09 +0000)]
Remove file accidentally added in r330320.

llvm-svn: 330321

6 years ago[IR/BasicBlockTest] Fix asan failure introduced in rL330316.
Florian Hahn [Thu, 19 Apr 2018 12:06:26 +0000 (12:06 +0000)]
[IR/BasicBlockTest] Fix asan failure introduced in rL330316.

The argument has to be deleted after the module containing the function
gets deleted.

llvm-svn: 330320

6 years ago[X86][FMA] Remove FMA reg-reg InstRW scheduler overrides.
Simon Pilgrim [Thu, 19 Apr 2018 11:37:26 +0000 (11:37 +0000)]
[X86][FMA] Remove FMA reg-reg InstRW scheduler overrides.

These are all already handled identically by WriteFMA.

llvm-svn: 330319

6 years ago[X86][BtVer2] Remove 128-bit F16C InstRW overrides.
Simon Pilgrim [Thu, 19 Apr 2018 11:16:33 +0000 (11:16 +0000)]
[X86][BtVer2] Remove 128-bit F16C InstRW overrides.

These are already handled identically by WriteCvtF2F.

llvm-svn: 330318

6 years ago[llvm-exegesis] Fix PfmIssueCountersTable creation
Simon Pilgrim [Thu, 19 Apr 2018 10:59:49 +0000 (10:59 +0000)]
[llvm-exegesis] Fix PfmIssueCountersTable creation

This patch ensures that the pfm issue counter tables are the correct size, accounting for the invalid resource entry at the beginning of the resource tables.

It also fixes an issue with pfm failing to match event counters due to a trailing comma added to all the event names.

I've also added a counter comment to each entry as it helps locate problems with the tables.

Note: I don't have access to a SandyBridge test machine, which is the only model to make use of multiple event counters being mapped to a single resource. I don't know if pfm accepts a comma-seperated list or not, but that is what it was doing.

Differential Revision: https://reviews.llvm.org/D45787

llvm-svn: 330317

6 years ago[BasicBlock] Add instructionsWithoutDebug methods to skip debug insts.
Florian Hahn [Thu, 19 Apr 2018 09:48:07 +0000 (09:48 +0000)]
[BasicBlock] Add instructionsWithoutDebug methods to skip debug insts.

Reviewers: aprantl, vsk, mattd, chandlerc

Reviewed By: aprantl, vsk

Differential Revision: https://reviews.llvm.org/D45657

llvm-svn: 330316

6 years ago[mips] Guard some macro expansions properly
Simon Dardis [Thu, 19 Apr 2018 09:45:04 +0000 (09:45 +0000)]
[mips] Guard some macro expansions properly

Reviewers: atanasyan, abeserminji

Differential Revision: https://reviews.llvm.org/D45565

llvm-svn: 330315

6 years agoAttempt to fix TestMiniDump on windows
Pavel Labath [Thu, 19 Apr 2018 09:38:42 +0000 (09:38 +0000)]
Attempt to fix TestMiniDump on windows

It was failing because the modules names were coming out as
C:\Windows\System32/MSVCP120D.dll (last separator is a forward slash) on
windows.

There are two issues at play here:
- the first problem is that the paths in minidump were being parsed as a
  host path. This meant that on posix systems the whole path was
  interpreted as a file name.
- on windows the path was split into a directory-filename pair
  correctly, but then when it was reconsituted, the last separator ended
  up being a forward slash because SBFileSpec.fullpath was joining them
  with '/' unconditionally.

I fix the first issue by parsing the minidump paths according to the
path syntax of the host which produced the dump, which should make the
test behavior on posix&windows identical. The last path will still be a
forward slash because of the second issue. We should probably fix the
"fullpath" property to do something smarter in the future.

llvm-svn: 330314

6 years ago[ARM] Add some missing FP16 VSEL test cases
Sjoerd Meijer [Thu, 19 Apr 2018 08:21:50 +0000 (08:21 +0000)]
[ARM] Add some missing FP16 VSEL test cases

Differential Revision: https://reviews.llvm.org/D45724

llvm-svn: 330313

6 years agotsan: fix compiler warnings
Dmitry Vyukov [Thu, 19 Apr 2018 07:42:08 +0000 (07:42 +0000)]
tsan: fix compiler warnings

vmaSize is uptr, so we need to print it with %zd.

llvm-svn: 330312

6 years ago[AArch64][AsmParser] NFC: Cleanup parsing of scalar registers.
Sander de Smalen [Thu, 19 Apr 2018 07:35:08 +0000 (07:35 +0000)]
[AArch64][AsmParser] NFC: Cleanup parsing of scalar registers.

Summary:
- Renamed tryParseRegister to tryParseScalarRegister, which
  now returns an OperandMatchResultTy.
- Moved matching of certain aliases into matchRegisterNameAlias.
- Changed type of most 'Reg' variables to 'unsigned'.

This is patch [1/4] in a series to add assembler/disassembler support for
SVE's contiguous LD1 (scalar+scalar) instructions:
- Patch [1/4]: https://reviews.llvm.org/D45687
- Patch [2/4]: https://reviews.llvm.org/D45688
- Patch [3/4]: https://reviews.llvm.org/D45689
- Patch [4/4]: https://reviews.llvm.org/D45690

Reviewers: fhahn, rengolin, javed.absar, huntergr, SjoerdMeijer, t.p.northover, echristo, evandro, samparker

Reviewed By: samparker

Subscribers: samparker, llvm-commits, kristof.beyls

Differential Revision: https://reviews.llvm.org/D45687

llvm-svn: 330311

6 years agoOpenBSD add C++ runtime in a driver's standpoint
Dean Michael Berris [Thu, 19 Apr 2018 06:55:30 +0000 (06:55 +0000)]
OpenBSD add C++ runtime in a driver's standpoint

Summary: - Since 6.2 release, on supporters platforms clang is shipped with both libcxx and libcxxabi.

Reviewers: dberris, alekseyshl, EricWF

Reviewed By: dberris

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D45662

llvm-svn: 330310

6 years ago[RuntimeDebugBuilder] Do not break for 64 bit integers
Tobias Grosser [Thu, 19 Apr 2018 05:38:12 +0000 (05:38 +0000)]
[RuntimeDebugBuilder] Do not break for 64 bit integers

In r330292 this assert was turned incorrectly into an unreachable, but
the correct behavior (thanks Michael) is to assert for anything that is
not 64 bit, but falltrough for 64 bit. I document this in the source
code.

llvm-svn: 330309

6 years ago[X86] Scrub scheduling information for MUL/IMUL on Intel CPUs.
Craig Topper [Thu, 19 Apr 2018 05:34:05 +0000 (05:34 +0000)]
[X86] Scrub scheduling information for MUL/IMUL on Intel CPUs.

This removes a bunch of unnecessary InstRW overrides. It also cleans up the missing information from the Sandy Bridge model. Other fixes to other models.

llvm-svn: 330308

6 years agoSimplify. NFC.
Rafael Espindola [Thu, 19 Apr 2018 03:51:26 +0000 (03:51 +0000)]
Simplify. NFC.

Using getOffset is here was a bit of an overkill. This is being
written and has relocations. This implies it is a .eh_frame or regular
section.

llvm-svn: 330307

6 years agoDon't call getOffset twice. NFC.
Rafael Espindola [Thu, 19 Apr 2018 02:24:28 +0000 (02:24 +0000)]
Don't call getOffset twice. NFC.

Just a bit faster.

llvm-svn: 330306

6 years agoFix test from r330245 on Windows.
Douglas Yung [Wed, 18 Apr 2018 23:58:05 +0000 (23:58 +0000)]
Fix test from r330245 on Windows.

llvm-svn: 330305

6 years ago[CodeGen] Do not push a destructor cleanup for a struct that doesn't
Akira Hatanaka [Wed, 18 Apr 2018 23:33:15 +0000 (23:33 +0000)]
[CodeGen] Do not push a destructor cleanup for a struct that doesn't
have a non-trivial destructor.

This fixes a bug introduced in r328731 where CodeGen emits calls to
synthesized destructors for non-trivial C structs in C++ mode when the
struct passed to EmitCallArg doesn't have a non-trivial destructor.
Under Microsoft's ABI, ASTContext::isParamDestroyedInCallee currently
always returns true, so it's necessary to check whether the struct has a
non-trivial destructor before pushing a cleanup in EmitCallArg.

This fixes PR37146.

llvm-svn: 330304

6 years ago[MS] Fix unprototyped thunk emission for incomplete return types
Reid Kleckner [Wed, 18 Apr 2018 23:21:32 +0000 (23:21 +0000)]
[MS] Fix unprototyped thunk emission for incomplete return types

Fixes PR37161

llvm-svn: 330303

6 years agoImprove LLDB's handling of non-local minidumps
Leonard Mosescu [Wed, 18 Apr 2018 23:10:46 +0000 (23:10 +0000)]
Improve LLDB's handling of non-local minidumps

Normally, LLDB is creating a high-fidelity representation of a live
process, including a list of modules and sections, with the
associated memory address ranges. In order to build the module and
section map LLDB tries to locate the local module image (object file)
and will parse it.

This does not work for postmortem debugging scenarios where the crash
dump (minidump in this case) was captured on a different machine.

Fortunately the minidump format encodes enough information about
each module's memory range to allow us to create placeholder modules.
This enables most LLDB functionality involving address-to-module
translations.

Also, we may want to completly disable the search for matching
local object files if we load minidumps unless we can prove that the
local image matches the one from the crash origin.
(not part of this change, see: llvm.org/pr35193)

Example: Identify the module from a stack frame PC:

Before:
  thread #1, stop reason = Exception 0xc0000005 encountered at address 0x164d14
    frame #0: 0x00164d14
    frame #1: 0x00167c79
    frame #2: 0x00167e6d
    frame #3: 0x7510336a
    frame #4: 0x77759882
    frame #5: 0x77759855

After:
  thread #1, stop reason = Exception 0xc0000005 encountered at address 0x164d14
    frame #0: 0x00164d14 C:\Users\amccarth\Documents\Visual Studio 2013\Projects\fizzbuzz\Debug\fizzbuzz.exe
    frame #1: 0x00167c79 C:\Users\amccarth\Documents\Visual Studio 2013\Projects\fizzbuzz\Debug\fizzbuzz.exe
    frame #2: 0x00167e6d C:\Users\amccarth\Documents\Visual Studio 2013\Projects\fizzbuzz\Debug\fizzbuzz.exe
    frame #3: 0x7510336a C:\Windows\SysWOW64\kernel32.dll
    frame #4: 0x77759882 C:\Windows\SysWOW64\ntdll.dll
    frame #5: 0x77759855 C:\Windows\SysWOW64\ntdll.dll

Example: target modules list

Before:
error: the target has no associated executable images

After:
[ 0] C:\Windows\System32\MSVCP120D.dll
[ 1] C:\Windows\SysWOW64\kernel32.dll
[ 2] C:\Users\amccarth\Documents\Visual Studio 2013\Projects\fizzbuzz\Debug\fizzbuzz.exe
[ 3] C:\Windows\System32\MSVCR120D.dll
[ 4] C:\Windows\SysWOW64\KERNELBASE.dll
[ 5] C:\Windows\SysWOW64\ntdll.dll

NOTE: the minidump format also includes the debug info GUID, so we can
fill-in the module UUID from it, but this part was excluded from this change
to keep the changes simple (the LLDB UUID is hardcoded to be either 16 or
20 bytes, while the CodeView GUIDs are normally 24 bytes)

Differential Revision: https://reviews.llvm.org/D45700

llvm-svn: 330302

6 years agoFix data race in X86FloatingPoint.cpp ASSERT_SORTED
Bob Haarman [Wed, 18 Apr 2018 23:04:09 +0000 (23:04 +0000)]
Fix data race in X86FloatingPoint.cpp ASSERT_SORTED

Summary:
ASSERT_SORTED checks if a table is sorted, and uses a boolean to
prevent the check from being run again if it was earlier determined
that the table is in fact sorted. Unsynchronized reads and writes of
that boolean triggered ThreadSanitizer's data race detection. This
change rewrites the code to use std::atomic<bool> instead.

Fixes PR36922.

Reviewers: rnk

Reviewed By: rnk

Subscribers: llvm-commits, hiraditya

Differential Revision: https://reviews.llvm.org/D45742

llvm-svn: 330301

6 years ago[COFF] Mark images with no exception handlers for /safeseh
Reid Kleckner [Wed, 18 Apr 2018 22:37:10 +0000 (22:37 +0000)]
[COFF] Mark images with no exception handlers for /safeseh

Summary:
DLLs and executables with no exception handlers need to be marked with
IMAGE_DLL_CHARACTERISTICS_NO_SEH, even if they have a load config.

Discovered here when building Chromium with LLD on Windows:
https://crbug.com/833951

Reviewers: ruiu, mstorsjo

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D45778

llvm-svn: 330300

6 years ago[OpenMP] Compilation error fix on const char*
Heejin Ahn [Wed, 18 Apr 2018 22:23:31 +0000 (22:23 +0000)]
[OpenMP] Compilation error fix on const char*

Summary:
This line
(https://github.com/llvm-mirror/openmp/blob/0ed912c7a798f5c4f65f8bb6b492e07fab7f4cea/runtime/src/kmp_gsupport.cpp#L1459)
added in D45327 (rL330282) causes a compilation failure.

Reviewers: jlpeyton

Subscribers: guansong, openmp-commits

Differential Revision: https://reviews.llvm.org/D45786

llvm-svn: 330299

6 years ago[X86] Correct the Defs, Uses, hasSideEffects, mayLoad, mayStore for XCHG and XADD...
Craig Topper [Wed, 18 Apr 2018 22:07:53 +0000 (22:07 +0000)]
[X86] Correct the Defs, Uses, hasSideEffects, mayLoad, mayStore for XCHG and XADD instructions.

I don't think we emit any of these from codegen except for using XCHG16ar as 2 byte NOP.

llvm-svn: 330298

6 years ago[HWASan] Add "N" suffix to generic __hwasan_load/store.
Alex Shlyapnikov [Wed, 18 Apr 2018 22:05:18 +0000 (22:05 +0000)]
[HWASan] Add "N" suffix to generic __hwasan_load/store.

Summary:
"N" suffix is added by the instrumentation and interface functions
are expected to be exported from the library as __hwasan_loadN* and
__hwasan_storeN*.

Reviewers: eugenis

Subscribers: kubamracek, delcypher, #sanitizers, llvm-commits

Differential Revision: https://reviews.llvm.org/D45739

llvm-svn: 330297

6 years ago[NVPTX, CUDA] Added support for m8n32k16 and m32n8k16 variants of wmma instructions.
Artem Belevich [Wed, 18 Apr 2018 21:51:48 +0000 (21:51 +0000)]
[NVPTX, CUDA] Added support for m8n32k16 and m32n8k16 variants of wmma instructions.

The new instructions were added added for sm_70+ GPUs in CUDA-9.1.

Differential Revision: https://reviews.llvm.org/D45068

llvm-svn: 330296

6 years ago[llvm-mca][X86] Add mmx versions of SSSE3 instructions
Simon Pilgrim [Wed, 18 Apr 2018 20:47:48 +0000 (20:47 +0000)]
[llvm-mca][X86] Add mmx versions of SSSE3 instructions

Move PABS instructions incorrectly tested under SSE2

llvm-svn: 330295

6 years ago[RISCV] Add test changes missed from rL330293
Alex Bradbury [Wed, 18 Apr 2018 20:36:12 +0000 (20:36 +0000)]
[RISCV] Add test changes missed from rL330293

llvm-svn: 330294

6 years ago[RISCV] Introduce pattern for materialising immediates with 0 for lower 12 bits
Alex Bradbury [Wed, 18 Apr 2018 20:34:23 +0000 (20:34 +0000)]
[RISCV] Introduce pattern for materialising immediates with 0 for lower 12 bits

These immediates can be materialised with just an lui, rather than an lui+addi
pair.

llvm-svn: 330293

6 years ago[RuntimeDebugBuilder] Print vectors passed without withspaces
Tobias Grosser [Wed, 18 Apr 2018 20:28:26 +0000 (20:28 +0000)]
[RuntimeDebugBuilder] Print vectors passed without withspaces

Originally the RuntimeDebugBuilder printed vectors with withspaces
between the elements. This historic use is meanwhile gone, but the
functionality is still available.

We now change the behavior to print elements just one after the other
without adding white spaces in between. This is useful for D45743, an
upcoming commmit, which also adds test coverage for this feature.

In general, printing elements of a vector directly is more generic as
it allows uses where no white-spaces are desired. Specifically, it
allows the user to build vectors of items to be printed where their
length is only known at run-time.

llvm-svn: 330292

6 years ago[RISCV] Add imm-cse.ll test case
Alex Bradbury [Wed, 18 Apr 2018 20:25:07 +0000 (20:25 +0000)]
[RISCV] Add imm-cse.ll test case

This test case demonstrates that common subexpression elimination takes place
between code sequences for materialising constants. In particular, it
demonstrates that redundant lui aren't generated. This would capture a
regression if applying a patch such as D41949.

llvm-svn: 330291

6 years ago[NFC] test case clean up
Lei Huang [Wed, 18 Apr 2018 20:22:26 +0000 (20:22 +0000)]
[NFC] test case clean up

1. remove redundant tests
2. update XForm_tests to generated expected code gen

llvm-svn: 330290

6 years ago[RuntimeDebugBuilder] Turn assert into an unreachable
Tobias Grosser [Wed, 18 Apr 2018 20:18:43 +0000 (20:18 +0000)]
[RuntimeDebugBuilder] Turn assert into an unreachable

llvm-svn: 330289

6 years ago[RISCV] Expand codegen -> compression sanity checks and move to a single file
Alex Bradbury [Wed, 18 Apr 2018 20:17:29 +0000 (20:17 +0000)]
[RISCV] Expand codegen -> compression sanity checks and move to a single file

The objdump tests interfere with update_llc_test_checks.py and can't be
automatically update them. Put the sanitify check for compression on the
codegen codepath into a separate file, and expand it to also include tests of
integer materialisation. This would catch changes such as those triggered by
D41949.

llvm-svn: 330288

6 years ago[X86] Fix the Uses/Defs,mayLoad,mayStore,hasSideEffects flags for the CMPXCHG instruc...
Craig Topper [Wed, 18 Apr 2018 20:15:00 +0000 (20:15 +0000)]
[X86] Fix the Uses/Defs,mayLoad,mayStore,hasSideEffects flags for the CMPXCHG instructions.

The compiler only emits the locked version of these which use different instruction definitions. The versions fixed here are only used by the assembler/disassembler.

llvm-svn: 330287

6 years agoadd extra acronyms for objc property names
Yan Zhang [Wed, 18 Apr 2018 20:09:10 +0000 (20:09 +0000)]
add extra acronyms for objc property names

Summary: This is to support general acronyms in Objective-C like 2G/3G/4G/... and coordinates X, Y, Z and W.

Reviewers: benhamilton

Reviewed By: benhamilton

Subscribers: klimek, cfe-commits

Differential Revision: https://reviews.llvm.org/D45750

llvm-svn: 330286

6 years ago[ScopDetect / ScopInfo] Get statistics for scops without any loop correctly
Tobias Grosser [Wed, 18 Apr 2018 20:03:36 +0000 (20:03 +0000)]
[ScopDetect / ScopInfo] Get statistics for scops without any loop correctly

Make sure we also counts scops not containing any loops.

llvm-svn: 330285

6 years ago[OPENMP] Fix -Wunused-lambda-capture. NFC
Fangrui Song [Wed, 18 Apr 2018 19:32:01 +0000 (19:32 +0000)]
[OPENMP] Fix -Wunused-lambda-capture. NFC

llvm-svn: 330284

6 years ago[OpenMP] Fix affinity API for KMP_AFFINITY=none|compact|scatter
Jonathan Peyton [Wed, 18 Apr 2018 19:25:48 +0000 (19:25 +0000)]
[OpenMP] Fix affinity API for KMP_AFFINITY=none|compact|scatter

Currently, the affinity API reports garbage for the initial place list and any
thread's place lists when using KMP_AFFINITY=none|compact|scatter.
This patch does two things:

for KMP_AFFINITY=none, Creates a one entry table for the places, this way, the
initial place list is just a single place with all the proc ids in it. We also
set the initial place of any thread to 0 instead of KMP_PLACE_ALL so that the
thread reports that single place (place 0) instead of garbage (-1) when using
the affinity API.

When non-OMP_PROC_BIND affinity is used
(including KMP_AFFINITY=compact|scatter), a thread's place list is populated
correctly. We assume that each thread is assigned to a single place. This is
implemented in two of the affinity API functions

Differential Revision: https://reviews.llvm.org/D45527

llvm-svn: 330283

6 years agoIntroduce GOMP_taskloop API
Jonathan Peyton [Wed, 18 Apr 2018 19:23:54 +0000 (19:23 +0000)]
Introduce GOMP_taskloop API

This patch introduces GOMP_taskloop to our API. It adds GOMP_4.5 to our
version symbols. Being a wrapper around __kmpc_taskloop, the function
creates a task with the loop bounds properly nested in the shareds so that
the GOMP task thunk will work properly. Also, the firstprivate copy constructors
are properly handled using the __kmp_gomp_task_dup() auxiliary function.

Currently, only linear spawning of tasks is supported
for the GOMP_taskloop interface.

Differential Revision: https://reviews.llvm.org/D45327

llvm-svn: 330282

6 years agoRevert "[RISCV] implement li pseudo instruction"
Alex Bradbury [Wed, 18 Apr 2018 19:02:31 +0000 (19:02 +0000)]
Revert "[RISCV] implement li pseudo instruction"

Reverts rL330224, while issues with the C extension and missed common
subexpression elimination opportunities are addressed. Neither of these issues
are visible in current RISC-V backend unit tests, which clearly need
expanding.

llvm-svn: 330281

6 years ago[CUDA] added missing __ldg(const signed char *)
Artem Belevich [Wed, 18 Apr 2018 18:33:43 +0000 (18:33 +0000)]
[CUDA] added missing __ldg(const signed char *)

Differential Revision: https://reviews.llvm.org/D45780

llvm-svn: 330280

6 years ago[HIP] Add driver input type for HIP
Yaxun Liu [Wed, 18 Apr 2018 18:25:03 +0000 (18:25 +0000)]
[HIP] Add driver input type for HIP

Patch by Greg Rodgers.
Revised by Yaxun Liu.

Differential Revision: https://reviews.llvm.org/D45489

llvm-svn: 330279

6 years ago[Power9]Legalize and emit code for converting Unsigned HWord/Char to Quad-Precision
Lei Huang [Wed, 18 Apr 2018 17:41:46 +0000 (17:41 +0000)]
[Power9]Legalize and emit code for converting Unsigned HWord/Char to Quad-Precision

Legalize and emit code for converting unsigned HWord/Char to QP:

xscvsdqp
xscvudqp

Only covering patterns for unsigned forms cause we don't have part-word
sign-extending integer loads into VSX registers.

Differential Revision: https://reviews.llvm.org/D45494

llvm-svn: 330278

6 years ago[MinGW] Try to fix asan testing after r330244
Martin Storsjo [Wed, 18 Apr 2018 17:34:29 +0000 (17:34 +0000)]
[MinGW] Try to fix asan testing after r330244

Twines shouldn't be stored as they can refer to temporaries.

llvm-svn: 330277

6 years ago[AArch64] Add isel pattern for v8i8->v2f32 NVCASTs.
Amara Emerson [Wed, 18 Apr 2018 17:10:19 +0000 (17:10 +0000)]
[AArch64] Add isel pattern for v8i8->v2f32 NVCASTs.

rdar://39454635

llvm-svn: 330276

6 years ago[LIT] Have lit run the lldb test suite
Jonas Devlieghere [Wed, 18 Apr 2018 17:08:49 +0000 (17:08 +0000)]
[LIT] Have lit run the lldb test suite

This is the first in what will hopefully become a series of patches to
replace the driver logic in dotest.py with LIT. The motivation for this
change is that there's no point in maintaining two driver
implementations. Since all of the LLVM projects are using lit, this is
the obvious choice.

Obviously the goal is maintain full compatibility with the functionality
offered by dotest. As such we won't be removing anything until that
point has been reached.

This patch is the initial attempt (referred to as v1) to run the lldb
test suite with lit. To do so we introduced a custom LLDB test format
that invokes dotest.py with a single test file.

Differential revision: https://reviews.llvm.org/D45333

llvm-svn: 330275

6 years ago[RISCV] Add specific tests for materialising imm32hi20 constants
Alex Bradbury [Wed, 18 Apr 2018 16:43:03 +0000 (16:43 +0000)]
[RISCV] Add specific tests for materialising imm32hi20 constants

i.e. constants that can be materialised with a single lui, as the lower 12
bits are zero.

llvm-svn: 330274

6 years ago[Power9]Legalize and emit code for converting (Un)Signed Word to Quad-Precision
Lei Huang [Wed, 18 Apr 2018 16:34:22 +0000 (16:34 +0000)]
[Power9]Legalize and emit code for converting (Un)Signed Word to Quad-Precision

Legalize and emit code for converting (Un)Signed Word to quad-precision via:

xscvsdqp
xscvudqp

Differential Revision: https://reviews.llvm.org/D45389

llvm-svn: 330273

6 years ago[NVPTX] Emit debug info in DWARF-2 by default for Cuda devices.
Alexey Bataev [Wed, 18 Apr 2018 16:31:09 +0000 (16:31 +0000)]
[NVPTX] Emit debug info in DWARF-2 by default for Cuda devices.

Summary:
NVPTX target supports debug info in DWARF-2 format. Patch adds emission
of debug info in DWARF-2 by default.

Reviewers: tra, jlebar

Subscribers: aprantl, JDevlieghere, cfe-commits

Differential Revision: https://reviews.llvm.org/D42581

llvm-svn: 330272

6 years ago[DEBUG] Initial adaptation of NVPTX target for debug info emission.
Alexey Bataev [Wed, 18 Apr 2018 16:13:41 +0000 (16:13 +0000)]
[DEBUG] Initial adaptation of NVPTX target for debug info emission.

Summary:
Patch adds initial emission of the debug info for NVPTX target.
Currently, only .file and .loc directives are emitted, everything else is
commented out to not break the compilation of Cuda.

Reviewers: echristo, jlebar, tra, jholewinski

Subscribers: mgorny, aprantl, JDevlieghere, llvm-commits

Differential Revision: https://reviews.llvm.org/D41827

llvm-svn: 330271

6 years ago[OPENMP] Code cleanup and code improvements.
Alexey Bataev [Wed, 18 Apr 2018 15:57:46 +0000 (15:57 +0000)]
[OPENMP] Code cleanup and code improvements.

llvm-svn: 330270

6 years ago[x86] Switch EFLAGS copy lowering to use reg-reg form of testing for
Chandler Carruth [Wed, 18 Apr 2018 15:52:50 +0000 (15:52 +0000)]
[x86] Switch EFLAGS copy lowering to use reg-reg form of testing for
a zero register.

Previously I tried this and saw LLVM unable to transform this to fold
with memory operands such as spill slot rematerialization. However, it
clearly works as shown in this patch. We turn these into `cmpb $0,
<mem>` when useful for folding a memory operand without issue. This form
has no disadvantage compared to `testb $-1, <mem>`. So overall, this is
likely no worse and may be slightly smaller in some cases due to the
`testb %reg, %reg` form.

Differential Revision: https://reviews.llvm.org/D45475

llvm-svn: 330269

6 years ago[sanitizer] Minor refactor of ThreadDescriptorSize
Kostya Kortchinsky [Wed, 18 Apr 2018 15:30:08 +0000 (15:30 +0000)]
[sanitizer] Minor refactor of ThreadDescriptorSize

Summary:
While I was sifting through dead code findings, I stumbled on this function.

First, for `__s390__` it always returned 0 for the 1st call, which doesn't seem
right. 2nd call & beyond would return the correct value though.
Then it duplicated the `atomic_store` multiple times, sometimes with a `if`,
sometimes without. Finally it used a capitalized variable name starting with `k`
which indicates a constant, and it is not.

So:
- rename the static global variable;
- change the atomic functions to their relaxed version;
- move the store to the end, and make sure we return `val` all the time.

Reviewers: alekseyshl, eugenis, koriakin

Reviewed By: alekseyshl

Subscribers: kubamracek, delcypher, llvm-commits, #sanitizers

Differential Revision: https://reviews.llvm.org/D45725

llvm-svn: 330268

6 years ago[llvm-mca] Use WithColor for printing errors
Jonas Devlieghere [Wed, 18 Apr 2018 15:26:51 +0000 (15:26 +0000)]
[llvm-mca] Use WithColor for printing errors

Use convenience helpers in WithColor to print errors and notes.

Differential revision: https://reviews.llvm.org/D45666

llvm-svn: 330267

6 years ago[support] Revert the changes made to Path.inc for the default Windows code page
Aaron Smith [Wed, 18 Apr 2018 15:26:26 +0000 (15:26 +0000)]
[support] Revert the changes made to Path.inc for the default Windows code page

Path.inc/widenPath tries to decode the path using both UTF-8 and the default Windows code page.
This is no longer necessary with the new InitLLVM method which ensures that the command line
arguemnts are already UTF-8 on Windows.

llvm-svn: 330266

6 years agoFix macosx build broken by r330249
Pavel Labath [Wed, 18 Apr 2018 15:23:21 +0000 (15:23 +0000)]
Fix macosx build broken by r330249

It seems llc crashes when targetting darwin with split-dwarf (pr37164).
This happens on all inputs, not just the one I added in the above
commit. Work around the issue by hardcoding the target triple to linux,
which is what all split-dwarf tests seem to be doing.

As I don't know of a way to specify the os part of the triple without
spelling out the architecture as well, I move the new test to the X86
folder.

llvm-svn: 330265

6 years ago[x86] Fix PR37100 by teaching the EFLAGS copy lowering to rewrite uses
Chandler Carruth [Wed, 18 Apr 2018 15:13:16 +0000 (15:13 +0000)]
[x86] Fix PR37100 by teaching the EFLAGS copy lowering to rewrite uses
across basic blocks in the limited cases where it is very straight
forward to do so.

This will also be useful for other places where we do some limited
EFLAGS propagation across CFG edges and need to handle copy rewrites
afterward. I think this is rapidly approaching the maximum we can and
should be doing here. Everything else begins to require either heroic
analysis to prove how to do PHI insertion manually, or somehow managing
arbitrary PHI-ing of EFLAGS with general PHI insertion. Neither of these
seem at all promising so if those cases come up, we'll almost certainly
need to rewrite the parts of LLVM that produce those patterns.

We do now require dominator trees in order to reliably diagnose patterns
that would require PHI nodes. This is a bit unfortunate but it seems
better than the completely mysterious crash we would get otherwise.

Differential Revision: https://reviews.llvm.org/D45673

llvm-svn: 330264

6 years ago[llvm-exegesis] Early out if the scheduler models have no extra info.
Simon Pilgrim [Wed, 18 Apr 2018 14:46:54 +0000 (14:46 +0000)]
[llvm-exegesis] Early out if the scheduler models have no extra info.

We were calling getExtraProcessorInfo() without checking hasExtraProcessorInfo(), resulting in an assertion.

llvm-svn: 330263

6 years ago[llvm-profdata] Use WithColor for printing errors
Jonas Devlieghere [Wed, 18 Apr 2018 14:42:33 +0000 (14:42 +0000)]
[llvm-profdata] Use WithColor for printing errors

Use convenience helpers in WithColor to print errors and warnings.

Differential revision: https://reviews.llvm.org/D45658

llvm-svn: 330262

6 years ago[llvm-link] Use WithColor for printing errors
Jonas Devlieghere [Wed, 18 Apr 2018 14:41:47 +0000 (14:41 +0000)]
[llvm-link] Use WithColor for printing errors

Use convenience helpers in WithColor to print errors and warnings.

Differential revision: https://reviews.llvm.org/D45667

llvm-svn: 330261

6 years ago[llvm-exegesis] Use LLVMTargetMachine pointer everywhere. NFCI.
Simon Pilgrim [Wed, 18 Apr 2018 14:22:33 +0000 (14:22 +0000)]
[llvm-exegesis] Use LLVMTargetMachine pointer everywhere. NFCI.

Avoid calling the unique_ptr multiple times.

llvm-svn: 330260

6 years ago[SimplifyLibcalls] Realloc(null, N) -> Malloc(N)
Sanjay Patel [Wed, 18 Apr 2018 14:21:31 +0000 (14:21 +0000)]
[SimplifyLibcalls] Realloc(null, N) -> Malloc(N)

Patch by Dávid Bolvanský!

Differential Revision: https://reviews.llvm.org/D45413

llvm-svn: 330259

6 years ago[llvm-exegesis] Put a newline at the end of each error report.
Simon Pilgrim [Wed, 18 Apr 2018 13:58:41 +0000 (13:58 +0000)]
[llvm-exegesis] Put a newline at the end of each error report.

Makes multiple error messages much easier to read.

llvm-svn: 330258

6 years ago[AMDGPU] Fix issues for backend divergence tracking
David Stuttard [Wed, 18 Apr 2018 13:53:31 +0000 (13:53 +0000)]
[AMDGPU] Fix issues for backend divergence tracking

Summary:
A change to use divergence analysis in the AMDGPU backend was getting formal
arguments incorrect (not tagged as divergent) unless they were VGPR0, VGPR1 or
VGPR2

For graphics shaders it is possible to have more than these passed in as VGPR

Modified the checking code to check for any VGPR registers passed in as formal
arguments.

Also, some intrinsics that are sources of divergence may have been lowered
during instruction selection and are missed on subsequent calls to
isSDNodeSourceOfDivergence - added the relevant AMDGPUISD checks as well.

Finally, the FunctionLoweringInfo tracks virtual registers that are live across
basic block boundaries. This is used to check for divergence of CopyFromRegister
registers using the DivergenceAnalysis analysis. For multiple blocks the lazily
evaluated inverted map VirtReg2Value was not cleared when the ValueMap map was.

Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D45372

Change-Id: I112f3bd6dfe0f62e63ce9b43b893982778e4bee3
llvm-svn: 330257

6 years ago[IRCE] Only check for NSW on equality predicates
Sam Parker [Wed, 18 Apr 2018 13:50:28 +0000 (13:50 +0000)]
[IRCE] Only check for NSW on equality predicates

After investigation discussed in D45439, it would seem that the nsw
flag restriction is unnecessary in most cases. So the IsInductionVar
lambda has been removed, the functionality extracted, and now only
require nsw when using eq/ne predicates.

Differential Revision: https://reviews.llvm.org/D45617

llvm-svn: 330256

6 years ago[llvm-exegesis] Pull out LLVMTargetMachine to simplify debugging. NFCI.
Simon Pilgrim [Wed, 18 Apr 2018 13:39:03 +0000 (13:39 +0000)]
[llvm-exegesis] Pull out LLVMTargetMachine to simplify debugging. NFCI.

Has been useful while trying to get around all the error reporting issues mentioned on PR37049.

llvm-svn: 330255

6 years ago[Sema] Disable built-in increment operator for bool in overload resolution in C++17
Jan Korous [Wed, 18 Apr 2018 13:38:39 +0000 (13:38 +0000)]
[Sema] Disable built-in increment operator for bool in overload resolution in C++17

Following: https://llvm.org/svn/llvm-project/cfe/trunk@329804

For C++17 the wording of [over.built] p4 excluded bool:

For every pair (T , vq), where T is an arithmetic type other than bool, there exist
candidate operator functions of the form
  vq T & operator++(vq T &);
  T operator++(vq T &, int);

Differential Revision: https://reviews.llvm.org/D45569

llvm-svn: 330254

6 years agoAdd tests for shrink wrapping and VLAs
Momchil Velikov [Wed, 18 Apr 2018 13:37:12 +0000 (13:37 +0000)]
Add tests for shrink wrapping and VLAs

Differential revision: https://reviews.llvm.org/D45727

llvm-svn: 330253

6 years ago[gold] Add support for optimization remarks
Teresa Johnson [Wed, 18 Apr 2018 13:25:23 +0000 (13:25 +0000)]
[gold] Add support for optimization remarks

Summary:
Adds support for LTO opt remarks (optionally with hotness) to
gold-plugin.

Reviewers: anemet

Subscribers: fhahn, mehdi_amini, llvm-commits

Differential Revision: https://reviews.llvm.org/D45752

llvm-svn: 330252

6 years ago[cmake] Improve pthread_[gs]etname_np detection code
Pavel Labath [Wed, 18 Apr 2018 13:13:27 +0000 (13:13 +0000)]
[cmake] Improve pthread_[gs]etname_np detection code

Summary:
Due to some android peculiarities, in some build configurations
(statically linked executables targeting older releases) we could detect
the presence of these functions (because they are present in libc.a,
where check_library_exists searches), but then fail to build because the
headers did not include the definition.

This attempts to remedy that by upgrading the check_library_exists to
check_symbol_exists, which will check that the function is declared too.

I am hoping that a more thorough check will make the messy #ifdef we
have accumulated in the code obsolete, so I optimistically try to remove
them.

Reviewers: zturner, kparzysz, danalbert

Subscribers: srhines, mgorny, krytarowski, llvm-commits

Differential Revision: https://reviews.llvm.org/D45359

llvm-svn: 330251

6 years ago[LoopUnroll] Only peel if a predicate becomes known in the loop body.
Florian Hahn [Wed, 18 Apr 2018 12:29:24 +0000 (12:29 +0000)]
[LoopUnroll] Only peel if a predicate becomes known in the loop body.

If a predicate does not become known after peeling, peeling is unlikely
to be beneficial.

Reviewers: mcrosier, efriedma, mkazantsev, junbuml

Reviewed By: mkazantsev

Differential Revision: https://reviews.llvm.org/D44983

llvm-svn: 330250

6 years ago[CodeGen/Dwarf] Make debug_names compatible with split-dwarf
Pavel Labath [Wed, 18 Apr 2018 12:11:59 +0000 (12:11 +0000)]
[CodeGen/Dwarf] Make debug_names compatible with split-dwarf

Summary:
Previously we crashed for the combination of the two features because we
tried to reference the dwo CU from the main object file. The fix
consists of two items:
- reference the skeleton CU from the name index (the consumer is
  expected to use the skeleton CU to find the real data).
- use the main object file string pool for the strings in the index

Reviewers: JDevlieghere, aprantl, dblaikie

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D45566

llvm-svn: 330249

6 years agoRevert r330195 "[NEON] Define vget_high_f16() and vget_low_f16() intrinsics in AArch6...
Ivan A. Kosarev [Wed, 18 Apr 2018 12:02:49 +0000 (12:02 +0000)]
Revert r330195 "[NEON] Define vget_high_f16() and vget_low_f16() intrinsics in AArch64 mode only".

Differential Revision: https://reviews.llvm.org/D45668

llvm-svn: 330248

6 years agoReport more precise error message when attach fails
Pavel Labath [Wed, 18 Apr 2018 11:56:21 +0000 (11:56 +0000)]
Report more precise error message when attach fails

Summary:
If the remote stub sends a specific error message instead of just a E??
code, we can use this to display a more informative error message
instead of just the generic "unable to attach" message.

I write a test for this using the SB API.
On the console this will show up like:
(lldb) process attach ...
error: attach failed: <STUB-MESSAGE>

if the stub supports error messages, or:
error: attach failed: Error ??

if it doesn't.

Reviewers: jingham, JDevlieghere

Subscribers: lldb-commits

Differential Revision: https://reviews.llvm.org/D45573

llvm-svn: 330247

6 years ago[UpdateTestChecks] Add update_mca_test_checks.py script
Greg Bedwell [Wed, 18 Apr 2018 10:27:45 +0000 (10:27 +0000)]
[UpdateTestChecks] Add update_mca_test_checks.py script

This script can be used to regenerate tests in the
test/tools/llvm-mca directory (PR36904).

Regenerated a number of tests using the pattern: test/tools/llvm-mca/*/*/*.s

Differential Revision: https://reviews.llvm.org/D45369

llvm-svn: 330246

6 years ago[clang-tidy] Fix clang-tidy doesn't read .clangtidy configuration file.
Haojian Wu [Wed, 18 Apr 2018 08:54:28 +0000 (08:54 +0000)]
[clang-tidy] Fix clang-tidy doesn't read .clangtidy configuration file.

Summary: Fix https://bugs.llvm.org/show_bug.cgi?id=34900.

Reviewers: alexfh

Reviewed By: alexfh

Subscribers: JonasToth, klimek, xazax.hun, cfe-commits

Differential Revision: https://reviews.llvm.org/D45697

llvm-svn: 330245

6 years ago[MinGW] Look for a cross sysroot relative to the clang binary
Martin Storsjo [Wed, 18 Apr 2018 08:47:26 +0000 (08:47 +0000)]
[MinGW] Look for a cross sysroot relative to the clang binary

If found, prefer this over looking for a similar gcc later in the
system path.

Differential Revision: https://reviews.llvm.org/D45504

llvm-svn: 330244

6 years ago[DebugInfo] Sink related dbg users when sinking in InstCombine
Bjorn Pettersson [Wed, 18 Apr 2018 08:08:04 +0000 (08:08 +0000)]
[DebugInfo] Sink related dbg users when sinking in InstCombine

Summary:
When sinking an instruction in InstCombine we now also sink
the DbgInfoIntrinsics that are using the sunken value.

Example)

When sinking the load in this input

bb.X:
  %0 = load i64, i64* %start, align 4, !dbg !31
  tail call void @llvm.dbg.value(metadata i64 %0, ...)
  br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
  br label %for.body

we now also move the dbg.value, like this

bb.X:
  br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
  %0 = load i64, i64* %start, align 4, !dbg !31
  tail call void @llvm.dbg.value(metadata i64 %0, ...)
  br label %for.body

In the past we haven't moved the dbg.value so we got

bb.X:
  tail call void @llvm.dbg.value(metadata i64 %0, ...)
  br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
  %0 = load i64, i64* %start, align 4, !dbg !31
  br label %for.body

So in the past we got a debug-use before the def of %0.
And that dbg.value was also on the path jumping to %for.end, for
which %0 never was defined.

CodeGenPrepare normally comes to rescue later (when not moving
the dbg.value), since it moves dbg.value instrinsics quite
brutally, without really analysing if it is correct to move
the intrinsic (see PR31878).
So at the moment this patch isn't expected to have much impact,
besides that it is moving the dbg.value already in opt, making
the IR look more sane directly.

This can be seen as a preparation to (hopefully) make it possible
to turn off CodeGenPrepare::placeDbgValues later as a solution
to PR31878.

I also adjusted test/DebugInfo/X86/sdagsplit-1.ll to make the
IR in the test case up-to-date with this behavior in InstCombine.

Reviewers: rnk, vsk, aprantl

Reviewed By: vsk, aprantl

Subscribers: mattd, JDevlieghere, llvm-commits

Tags: #debug-info

Differential Revision: https://reviews.llvm.org/D45425

llvm-svn: 330243

6 years ago[NFC] Remove doxygen brief tag from BasicBlock.h
Matt Davis [Wed, 18 Apr 2018 07:58:45 +0000 (07:58 +0000)]
[NFC] Remove doxygen brief tag from BasicBlock.h

Summary: Documentation is built using the AutoBrief configuration option in docs/doxygen.cfg.in.  This change removes the redundant brief tags from BasicBlock.h.   I'm happy to write a sed script and remove all \brief tags as a separate commit later.

Reviewers: aprantl

Reviewed By: aprantl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D45708

llvm-svn: 330242

6 years ago[X86][Broadwell] Remove some unnecessary InstRW overrides and add some FIXMEs.
Craig Topper [Wed, 18 Apr 2018 06:41:25 +0000 (06:41 +0000)]
[X86][Broadwell] Remove some unnecessary InstRW overrides and add some FIXMEs.

llvm-svn: 330241