platform/upstream/llvm.git
4 years ago[ExecutionEngine] Add JITSymbolFlags::fromSummary(GlobalValueSummary*)
Stefan Gränitz [Sat, 29 Feb 2020 11:52:19 +0000 (11:52 +0000)]
[ExecutionEngine] Add JITSymbolFlags::fromSummary(GlobalValueSummary*)

Summary: A function that creates JITSymbolFlags from a GlobalValueSummary. Similar functions exist: fromGlobalValue(), fromObjectSymbol()

Reviewers: lhames

Reviewed By: lhames

Subscribers: hiraditya, steven_wu, dexonsmith, arphaman, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75082

4 years agoFirst test commit - empty
Stefanos Baziotis [Sat, 29 Feb 2020 09:53:11 +0000 (11:53 +0200)]
First test commit - empty

4 years ago[llvm-readobj] - Report warnings instead of errors for broken relocations.
Georgii Rymar [Fri, 14 Feb 2020 09:47:52 +0000 (12:47 +0300)]
[llvm-readobj] - Report warnings instead of errors for broken relocations.

This is a follow-up for https://reviews.llvm.org/D74545.

It adds test cases for each incorrect case returned in `getRelocationTarget`.

Differential revision: https://reviews.llvm.org/D74595

4 years agoArrayRef'ize restoreCalleeSavedRegisters. NFCI.
Benjamin Kramer [Sat, 29 Feb 2020 08:50:23 +0000 (09:50 +0100)]
ArrayRef'ize restoreCalleeSavedRegisters. NFCI.

restoreCalleeSavedRegisters can mutate the contents of the
CalleeSavedInfos, so use a MutableArrayRef.

4 years ago[libc] Add ability to generate enum types/values to HdrGen.
Siva Chandra Reddy [Thu, 27 Feb 2020 22:30:24 +0000 (14:30 -0800)]
[libc] Add ability to generate enum types/values to HdrGen.

A target to generate the std C threads.h file has been added. This
utilizes the new feature added in this change.

Reviewers: phosek

Differential Revision: https://reviews.llvm.org/D75379

4 years ago[X86] Move the function getOrCreateBoundaryAlignFragment
Shengchen Kan [Fri, 28 Feb 2020 14:27:53 +0000 (22:27 +0800)]
[X86] Move the function getOrCreateBoundaryAlignFragment

MCObjectStreamer is more suitable to create fragments than
X86AsmBackend, for example, the function getOrCreateDataFragment is
defined in MCObjectStreamer.

Differential Revision: https://reviews.llvm.org/D75351

4 years ago[X86] Disable the NOP padding for branches when bundle is enabled
Shengchen Kan [Fri, 28 Feb 2020 13:09:30 +0000 (21:09 +0800)]
[X86] Disable the NOP padding for branches when bundle is enabled

When bundle is enabled, data fragment itself has a space to emit NOP
to bundle-align instructions. The behaviour makes it impossible for
us to determine whether the macro fusion really happen when emitting
instructions. In addition, boundary-align fragment is also used to
emit NOPs to align instructions, currently using them together sometimes
makes code crazy.

Differential Revision: https://reviews.llvm.org/D75346

4 years ago[ELF][PPC32] Fix canonical PLTs when the order does not match the PLT order
Fangrui Song [Sat, 29 Feb 2020 01:22:29 +0000 (17:22 -0800)]
[ELF][PPC32] Fix canonical PLTs when the order does not match the PLT order

Reviewed By: Bdragon28

Differential Revision: https://reviews.llvm.org/D75394

4 years ago[ELF] Delete two unneeded `referenced = true` after D65584
Fangrui Song [Sat, 29 Feb 2020 05:26:36 +0000 (21:26 -0800)]
[ELF] Delete two unneeded `referenced = true` after D65584

4 years agoRevert "Fix GSYM tests to run the yaml files and fix test failures on some machines."
Greg Clayton [Sat, 29 Feb 2020 05:19:05 +0000 (21:19 -0800)]
Revert "Fix GSYM tests to run the yaml files and fix test failures on some machines."

This reverts commit d334ce0b5acb945d6202d0ab6a17bdca530f50c1.

4 years ago[cmake] Fix LLVM_USE_SPLIT_DWARF
Michael Liao [Fri, 28 Feb 2020 05:01:08 +0000 (00:01 -0500)]
[cmake] Fix LLVM_USE_SPLIT_DWARF

Summary:
- Add `-gsplit-dwarf` as an option instead of a definition.
- Only add that option on compilers supporting dwarf splitting, such as clang
  and gcc.

Reviewers: echristo, pcc

Subscribers: mgorny, aprantl, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75328

4 years agoDisable instrprof-merging.cpp on powerpc64 (D69471 follow up)
Vedant Kumar [Sat, 29 Feb 2020 02:59:56 +0000 (18:59 -0800)]
Disable instrprof-merging.cpp on powerpc64 (D69471 follow up)

An execution count goes missing for a constructor, this needs
investigation:

http://lab.llvm.org:8011/builders/clang-ppc64be-linux/builds/45132/

```
/home/buildbots/ppc64be-clang-test/clang-ppc64be/llvm/compiler-rt/test/profile/instrprof-merging.cpp:28:16:
error: V1: expected string not found in input
 A() {} // V1: [[@LINE]]{{ *}}|{{ *}}1

<stdin>:28:32: note: possible intended match here
 28| | A() {} // V1: [[@LINE]]{{ *}}|{{ *}}1
```

4 years agoAdd cast to appease clang-armv7-linux-build-cache (D69471 followup)
Vedant Kumar [Sat, 29 Feb 2020 02:27:04 +0000 (18:27 -0800)]
Add cast to appease clang-armv7-linux-build-cache (D69471 followup)

http://lab.llvm.org:8011/builders/clang-armv7-linux-build-cache/builds/27075

error: non-constant-expression cannot be narrowed from type 'uint64_t'
(aka 'unsigned long long') to 'size_t' (aka 'unsigned int') in
initializer list [-Wc++11-narrowing]
  return {MappingBuf, getDataSize<FuncRecordTy, Endian>(Record)};

4 years agoReland: [Coverage] Revise format to reduce binary size
Vedant Kumar [Mon, 21 Oct 2019 18:48:38 +0000 (11:48 -0700)]
Reland: [Coverage] Revise format to reduce binary size

Try again with an up-to-date version of D69471 (99317124 was a stale
revision).

---

Revise the coverage mapping format to reduce binary size by:

1. Naming function records and marking them `linkonce_odr`, and
2. Compressing filenames.

This shrinks the size of llc's coverage segment by 82% (334MB -> 62MB)
and speeds up end-to-end single-threaded report generation by 10%. For
reference the compressed name data in llc is 81MB (__llvm_prf_names).

Rationale for changes to the format:

- With the current format, most coverage function records are discarded.
  E.g., more than 97% of the records in llc are *duplicate* placeholders
  for functions visible-but-not-used in TUs. Placeholders *are* used to
  show under-covered functions, but duplicate placeholders waste space.

- We reached general consensus about giving (1) a try at the 2017 code
  coverage BoF [1]. The thinking was that using `linkonce_odr` to merge
  duplicates is simpler than alternatives like teaching build systems
  about a coverage-aware database/module/etc on the side.

- Revising the format is expensive due to the backwards compatibility
  requirement, so we might as well compress filenames while we're at it.
  This shrinks the encoded filenames in llc by 86% (12MB -> 1.6MB).

See CoverageMappingFormat.rst for the details on what exactly has
changed.

Fixes PR34533 [2], hopefully.

[1] http://lists.llvm.org/pipermail/llvm-dev/2017-October/118428.html
[2] https://bugs.llvm.org/show_bug.cgi?id=34533

Differential Revision: https://reviews.llvm.org/D69471

4 years agoRevert "[Coverage] Revise format to reduce binary size"
Vedant Kumar [Sat, 29 Feb 2020 02:03:15 +0000 (18:03 -0800)]
Revert "[Coverage] Revise format to reduce binary size"

This reverts commit 99317124e1c772e9a9de41a0cd56e1db049b4ea4. This is
still busted on Windows:

http://lab.llvm.org:8011/builders/lld-x86_64-win7/builds/40873

The llvm-cov tests report 'error: Could not load coverage information'.

4 years ago[compiler-rt/test] Relax a test so we can debug it on sanitizer-x86_64-linux-android
Vedant Kumar [Sat, 29 Feb 2020 01:46:54 +0000 (17:46 -0800)]
[compiler-rt/test] Relax a test so we can debug it on sanitizer-x86_64-linux-android

This test is failing with a core dump on /just/ this bot, and I'd like
to find out why.

http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-android/builds/27979/

4 years ago[Coverage] Revise format to reduce binary size
Vedant Kumar [Mon, 21 Oct 2019 18:48:38 +0000 (11:48 -0700)]
[Coverage] Revise format to reduce binary size

Revise the coverage mapping format to reduce binary size by:

1. Naming function records and marking them `linkonce_odr`, and
2. Compressing filenames.

This shrinks the size of llc's coverage segment by 82% (334MB -> 62MB)
and speeds up end-to-end single-threaded report generation by 10%. For
reference the compressed name data in llc is 81MB (__llvm_prf_names).

Rationale for changes to the format:

- With the current format, most coverage function records are discarded.
  E.g., more than 97% of the records in llc are *duplicate* placeholders
  for functions visible-but-not-used in TUs. Placeholders *are* used to
  show under-covered functions, but duplicate placeholders waste space.

- We reached general consensus about giving (1) a try at the 2017 code
  coverage BoF [1]. The thinking was that using `linkonce_odr` to merge
  duplicates is simpler than alternatives like teaching build systems
  about a coverage-aware database/module/etc on the side.

- Revising the format is expensive due to the backwards compatibility
  requirement, so we might as well compress filenames while we're at it.
  This shrinks the encoded filenames in llc by 86% (12MB -> 1.6MB).

See CoverageMappingFormat.rst for the details on what exactly has
changed.

Fixes PR34533 [2], hopefully.

[1] http://lists.llvm.org/pipermail/llvm-dev/2017-October/118428.html
[2] https://bugs.llvm.org/show_bug.cgi?id=34533

Differential Revision: https://reviews.llvm.org/D69471

4 years agoTry to fix WindowsManifest CMake logic on Windows
Reid Kleckner [Sat, 29 Feb 2020 01:21:50 +0000 (17:21 -0800)]
Try to fix WindowsManifest CMake logic on Windows

CMake is complaining about the "^" regex if the prefixes are empty
strings.

4 years agoFix GSYM tests to run the yaml files and fix test failures on some machines.
Greg Clayton [Fri, 28 Feb 2020 23:25:47 +0000 (15:25 -0800)]
Fix GSYM tests to run the yaml files and fix test failures on some machines.

Summary: YAML files were not being run during lit testing as there was no lit.local.cfg file. Once this was fixed, some buildbots would fail due to a StringRef that pointed to a std::string inside of a temporary llvm::Triple object. These issues are fixed here by making a local triple object that stays around long enough so the StringRef points to valid data.

Reviewers: aprantl, thakis, MaskRay, aadsm, wallace

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75390

4 years ago[X86] Remove isel patterns from broadcast of loadi32.
Craig Topper [Fri, 28 Feb 2020 23:42:47 +0000 (15:42 -0800)]
[X86] Remove isel patterns from broadcast of loadi32.

We already combine non extending loads with broadcasts in DAG
combine. All these patterns are picking up is the aligned extload
special case. But the only lit test we have that exercsises it is
using v8i1 load that datalayout is reporting align 8 for. That
seems generous. So without a realistic test case I don't think
there is much value in these patterns.

4 years ago[compiler-rt/test] Use FileCheck -allow-empty instead of count 0
Vedant Kumar [Sat, 29 Feb 2020 00:20:54 +0000 (16:20 -0800)]
[compiler-rt/test] Use FileCheck -allow-empty instead of count 0

Hope this fixes:

http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-android/builds/27977/steps/run%20lit%20tests%20%5Bi686%2Ffugu-userdebug%2FN2G48C%5D/logs/stdio

```
: 'RUN: at line 8';   UBSAN_OPTIONS=suppressions=/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/compiler_rt_build_android_i686/test/ubsan/Standalone-i386/TestCases/Misc/Output/nullability.c.tmp.supp  /var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/compiler_rt_build_android_i686/test/ubsan/Standalone-i386/TestCases/Misc/Output/nullability.c.tmp 2>&1 | count 0
--
Exit Code: 1

Command Output (stderr):
--
Expected 0 lines, got 2.
```

Not sure what this would be printing though, a sanitizer initialization message?

4 years ago[LTO][Legacy] Add explicit dependency on BinaryFormat
Francis Visoiu Mistrih [Fri, 28 Feb 2020 23:49:28 +0000 (15:49 -0800)]
[LTO][Legacy] Add explicit dependency on BinaryFormat

This fixes some windows bots.

4 years ago[DFSan] Add __dfsan_cmp_callback.
Matt Morehouse [Fri, 28 Feb 2020 23:49:37 +0000 (15:49 -0800)]
[DFSan] Add __dfsan_cmp_callback.

Summary:
When -dfsan-event-callbacks is specified, insert a call to
__dfsan_cmp_callback on every CMP instruction.

Reviewers: vitalybuka, pcc, kcc

Reviewed By: kcc

Subscribers: hiraditya, #sanitizers, eugenis, llvm-commits

Tags: #sanitizers, #llvm

Differential Revision: https://reviews.llvm.org/D75389

4 years ago[Sema] Fix an assert when objc_externally_retained was applied to an unprototyped...
Erik Pilkington [Fri, 28 Feb 2020 23:24:23 +0000 (15:24 -0800)]
[Sema] Fix an assert when objc_externally_retained was applied to an unprototyped function

rdar://58893199

4 years ago[DFSan] Add __dfsan_mem_transfer_callback.
Matt Morehouse [Fri, 28 Feb 2020 23:48:03 +0000 (15:48 -0800)]
[DFSan] Add __dfsan_mem_transfer_callback.

Summary:
When -dfsan-event-callbacks is specified, insert a call to
__dfsan_mem_transfer_callback on every memcpy and memmove.

Reviewers: vitalybuka, kcc, pcc

Reviewed By: kcc

Subscribers: eugenis, hiraditya, #sanitizers, llvm-commits

Tags: #sanitizers, #llvm

Differential Revision: https://reviews.llvm.org/D75386

4 years ago[modules] Allow frameworks to have only a private module without a public one.
Volodymyr Sapsai [Thu, 27 Feb 2020 23:51:24 +0000 (15:51 -0800)]
[modules] Allow frameworks to have only a private module without a public one.

Support only preferred spelling 'Modules/module.private.modulemap' and
not the deprecated 'module_private.map'.

rdar://problem/57715533

Reviewed By: bruno

Differential Revision: https://reviews.llvm.org/D75311

4 years ago[AMDGPU] Fix scheduling model for V_MULLIT_F32
Jay Foad [Fri, 28 Feb 2020 23:20:45 +0000 (23:20 +0000)]
[AMDGPU] Fix scheduling model for V_MULLIT_F32

This was incorrectly marked as a half rate 64-bit instruction by D45073.

4 years ago[X86] Canonicalize (bitcast (vbroadcast_load)) so that the cast and vbroadcast_load...
Craig Topper [Fri, 28 Feb 2020 08:35:52 +0000 (00:35 -0800)]
[X86] Canonicalize (bitcast (vbroadcast_load)) so that the cast and vbroadcast_load are both integer or fp.

Helps a little with some isel pattern matching. Especially on
32-bit targets where we sometimes use f64 loads.

4 years ago[X86] Remove stale FIXME form test. NFC.
Craig Topper [Fri, 28 Feb 2020 05:32:49 +0000 (21:32 -0800)]
[X86] Remove stale FIXME form test. NFC.

4 years ago[X86] Cleanup a comment around bitcasting X86ISD::VBROADCAST_LOAD and add an assert...
Craig Topper [Fri, 28 Feb 2020 05:29:52 +0000 (21:29 -0800)]
[X86] Cleanup a comment around bitcasting X86ISD::VBROADCAST_LOAD and add an assert to make sure memory VT size doesn't change.

4 years ago[llvm][Support][modulemap] Exclude WindowsSupport.h from the LLVM_Util module
Michael Spencer [Fri, 28 Feb 2020 22:39:49 +0000 (14:39 -0800)]
[llvm][Support][modulemap] Exclude WindowsSupport.h from the LLVM_Util module

rG01f9abbb50b1 moved WindowsSupport.h to include/llvm/Support/Windows/

This is a problem because the modulemap include all of the Support and
ADT directories, thus any use of any header in Support or ADT would
cause the compiler to try to build WindowsSupport.h, which only works
on Windows.

Fix this by explicitly excluding WindowsSupport.h from the LLVM_Util
module.

4 years ago[ubsan] Add support for -fsanitize=nullability-* suppressions
Vedant Kumar [Fri, 28 Feb 2020 22:09:14 +0000 (14:09 -0800)]
[ubsan] Add support for -fsanitize=nullability-* suppressions

rdar://59402904

4 years ago[entry values] ARM: Add a describeLoadedValue override (PR45025)
Vedant Kumar [Thu, 27 Feb 2020 17:58:24 +0000 (09:58 -0800)]
[entry values] ARM: Add a describeLoadedValue override (PR45025)

As a narrow stopgap for the assertion failure described in PR45025, add
a describeLoadedValue override to ARMBaseInstrInfo and use it to detect
copies in which the forwarding reg is a super/sub reg of the copy
destination. For the moment this is unsupported.

Several follow ups are possible:

1) Handle VORRq. At the moment, we do not, because isCopyInstrImpl
   returns early when !MI.isMoveReg().

2) In the case where forwarding reg is a super-reg of the copy
   destination, we should be able to describe the forwarding reg as a
   subreg within the copy destination. I'm not 100% sure about this, but
   it looks like that's what's done in AArch64InstrInfo.

3) In the case where the forwarding reg is a sub-reg of the copy
   destination, maybe we could describe the forwarding reg using the
   copy destinaion and a DW_OP_LLVM_fragment (I guess this should be
   possible after D75036).

https://bugs.llvm.org/show_bug.cgi?id=45025
rdar://59772698

Differential Revision: https://reviews.llvm.org/D75273

4 years ago[DFSan] Add __dfsan_load_callback.
Matt Morehouse [Fri, 28 Feb 2020 22:25:45 +0000 (14:25 -0800)]
[DFSan] Add __dfsan_load_callback.

Summary:
When -dfsan-event-callbacks is specified, insert a call to
__dfsan_load_callback() on every load.

Reviewers: vitalybuka, pcc, kcc

Reviewed By: vitalybuka, kcc

Subscribers: hiraditya, #sanitizers, llvm-commits, eugenis, kcc

Tags: #sanitizers, #llvm

Differential Revision: https://reviews.llvm.org/D75363

4 years ago[ADT] Allow K to be incomplete during DenseMap<K*, V> instantiation
Reid Kleckner [Thu, 27 Feb 2020 21:11:17 +0000 (13:11 -0800)]
[ADT] Allow K to be incomplete during DenseMap<K*, V> instantiation

DenseMap requires two sentinel values for keys: empty and tombstone
values. To avoid undefined behavior, LLVM aligns the two sentinel
pointers to alignof(T). This requires T to be complete, which is
needlessly restrictive.

Instead, assume that DenseMap pointer keys have a maximum alignment of
4096, and use the same sentinel values for all pointer keys. The new
sentinels are:
  empty:     static_cast<uintptr_t>(-1) << 12
  tombstone: static_cast<uintptr_t>(-2) << 12

These correspond to the addresses of -4096 and -8192. Hopefully, such a
key is never inserted into a DenseMap.

I encountered this while looking at making clang's SourceManager not
require FileManager.h, but it has several maps keyed on classes defined
in FileManager.h. FileManager depends on various LLVM FS headers, which
cumulatively take ~200ms to parse, and are generally not needed.

Reviewed By: hans

Differential Revision: https://reviews.llvm.org/D75301

4 years ago[NFC] Fix minor python issues.
Jordan Rupprecht [Fri, 28 Feb 2020 22:11:21 +0000 (14:11 -0800)]
[NFC] Fix minor python issues.

* llvm-gisel-cov.py: extra `)` after print
* chunk-print-before-all.py: py2-only print method

4 years agoRevert "[MLIR] Move from using target_link_libraries to LINK_LIBS for llvm libraries."
Stephen Neuendorffer [Fri, 28 Feb 2020 22:13:30 +0000 (14:13 -0800)]
Revert "[MLIR] Move from using target_link_libraries to LINK_LIBS for llvm libraries."

This reverts commit 2f265e35285127de376afc2243e3b5d0c5842cc2.

4 years agoRevert "[MLIR] Remove redundant library dependencies"
Stephen Neuendorffer [Fri, 28 Feb 2020 22:04:58 +0000 (14:04 -0800)]
Revert "[MLIR] Remove redundant library dependencies"

This reverts commit e1cb15c8f92215585105fc00d6eab80860dbc68d.

4 years ago[AMDGPU] Remove dubious logic in bidirectional list scheduler
Jay Foad [Mon, 7 Oct 2019 14:33:59 +0000 (15:33 +0100)]
[AMDGPU] Remove dubious logic in bidirectional list scheduler

Summary:
pickNodeBidirectional tried to compare the best top candidate and the
best bottom candidate by examining TopCand.Reason and BotCand.Reason.
This is unsound because, after calling pickNodeFromQueue, Cand.Reason
does not reflect the most important reason why Cand was chosen. Rather
it reflects the most recent reason why it beat some other potential
candidate, which could have been for some low priority tie breaker
reason.

I have seen this cause problems where TopCand is a good candidate, but
because TopCand.Reason is ORDER (which is very low priority) it is
repeatedly ignored in favour of a mediocre BotCand. This is not how
bidirectional scheduling is supposed to work.

To fix this I changed the code to always compare TopCand and BotCand
directly, like the generic implementation of pickNodeBidirectional does.
This removes some uncommented AMDGPU-specific logic; if this logic turns
out to be important then perhaps it could be moved into an override of
tryCandidate instead.

Graphics shader benchmarking on gfx10 shows a lot more positive than
negative effects from this change.

Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB

Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68338

4 years ago[LLD] Report errors occurred while parsing debug info as warnings.
Alexey Lapshin [Thu, 27 Feb 2020 10:21:58 +0000 (13:21 +0300)]
[LLD] Report errors occurred while parsing debug info as warnings.

Summary:
Extracted from D74773. Currently, errors happened while parsing
debug info are reported as errors. DebugInfoDWARF library treats such
errors as "Recoverable errors". This patch makes debug info errors
to be reported as warnings, to support DebugInfoDWARF approach.

Reviewers: ruiu, grimar, MaskRay, jhenderson, espindola

Reviewed By: MaskRay, jhenderson

Subscribers: emaste, aprantl, arichardson, arphaman, llvm-commits

Tags: #llvm, #debug-info, #lld

Differential Revision: https://reviews.llvm.org/D75234

4 years ago[LTO][Legacy] Add new API to query Mach-O CPU (sub)type
Francis Visoiu Mistrih [Wed, 19 Feb 2020 01:13:47 +0000 (17:13 -0800)]
[LTO][Legacy] Add new API to query Mach-O CPU (sub)type

Tools working with object files on Darwin (e.g. lipo) may need to know
properties like the CPU type and subtype of a bitcode file. The logic of
converting a triple to a Mach-O CPU_(SUB_)TYPE should be provided by
LLVM instead of relying on tools to re-implement it.

Differential Revision: https://reviews.llvm.org/D75067

4 years ago[MLIR] Add vector support for fpexp and fptrunc.
Tim Shen [Tue, 25 Feb 2020 23:48:23 +0000 (15:48 -0800)]
[MLIR] Add vector support for fpexp and fptrunc.

Differential Revision: https://reviews.llvm.org/D75150

4 years ago[MLIR] Add llvm.mlir.cast op for semantic preserving cast between dialect types.
Tim Shen [Sat, 22 Feb 2020 02:13:56 +0000 (18:13 -0800)]
[MLIR] Add llvm.mlir.cast op for semantic preserving cast between dialect types.

Summary: See discussion here: https://llvm.discourse.group/t/rfc-dialect-type-cast-op/538/11

Reviewers: ftynse

Subscribers: bixia, sanjoy.google, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, Joonsoo, llvm-commits

Differential Revision: https://reviews.llvm.org/D75141

4 years ago[MLIR] Add LLVMConversionTarget as a customization point. NFC.
Tim Shen [Tue, 25 Feb 2020 21:31:56 +0000 (13:31 -0800)]
[MLIR] Add LLVMConversionTarget as a customization point. NFC.

This is in preparation for the next patch D75141. The purpose is to
provide a single place where LLVM dialect registers its ops as
legal/illegal.

Reviewers: ftynse, mravishankar, herhut

Subscribers: jholewinski, bixia, sanjoy.google, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, csigg, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, Joonsoo, llvm-commits

Differential Revision: https://reviews.llvm.org/D75140

4 years ago[Hexagon] Map dcfetch intrinsic to Y2_dcfetchbo, not Y2_dcfetch
Krzysztof Parzyszek [Fri, 28 Feb 2020 19:50:16 +0000 (13:50 -0600)]
[Hexagon] Map dcfetch intrinsic to Y2_dcfetchbo, not Y2_dcfetch

4 years agoRevert "[MLIR] Move from add_dependencies() to DEPENDS"
Stephen Neuendorffer [Fri, 28 Feb 2020 20:16:51 +0000 (12:16 -0800)]
Revert "[MLIR] Move from add_dependencies() to DEPENDS"

This reverts commit 8a2b86b2c20a50f4a32931ae311b8e2457e591eb.

4 years agoRevert "[MLIR] Add support for libMLIR.so"
Stephen Neuendorffer [Fri, 28 Feb 2020 20:16:10 +0000 (12:16 -0800)]
Revert "[MLIR] Add support for libMLIR.so"

This reverts commit 1246e867164b06fc3f0de6bfaaa0922d99cb5ce9.

4 years agoRevert "[MLIR] Fixes for BUILD_SHARED_LIBS=on"
Stephen Neuendorffer [Fri, 28 Feb 2020 20:15:45 +0000 (12:15 -0800)]
Revert "[MLIR] Fixes for BUILD_SHARED_LIBS=on"

This reverts commit c767dc9394b4cca26c3eb40f1270092594e539c2.

4 years ago[OpenMP5.0] Allow pointer arithmetic in motion/map clause, by Chi Chun
cchen [Fri, 28 Feb 2020 19:37:14 +0000 (14:37 -0500)]
[OpenMP5.0] Allow pointer arithmetic in motion/map clause, by Chi Chun
Chen

Summary:
Base declaration in pointer arithmetic expression is determined by
binary search with type information. Take "int *a, *b; *(a+*b)" as an
example, we determine the base by checking the type of LHS and RHS. In
this case the type of LHS is "int *", the type of RHS is "int",
therefore, we know that we need to visit LHS in order to find base
declaration.

Reviewers: ABataev, jdoerfert

Reviewed By: ABataev

Subscribers: guansong, cfe-commits, sandoval, dreachem

Tags: #clang

Differential Revision: https://reviews.llvm.org/D75077

4 years ago[MLIR] Fixes for BUILD_SHARED_LIBS=on
Stephen Neuendorffer [Thu, 27 Feb 2020 22:53:12 +0000 (14:53 -0800)]
[MLIR] Fixes for BUILD_SHARED_LIBS=on

Differential Revision: https://reviews.llvm.org/D75308

4 years ago[MLIR] Add support for libMLIR.so
Valentin Churavy [Sun, 9 Feb 2020 03:27:54 +0000 (19:27 -0800)]
[MLIR] Add support for libMLIR.so

Putting this up mainly for discussion on
how this should be done. I am interested in MLIR from
the Julia side and we currently have a strong preference
to dynamically linking against the LLVM shared library,
and would like to have a MLIR shared library.

This patch adds a new cmake function add_mlir_library()
which accumulates a list of targets to be compiled into
libMLIR.so.  Note that not all libraries make sense to
be compiled into libMLIR.so.  In particular, we want
to avoid libraries which primarily exist to support
certain tools (such as mlir-opt and mlir-cpu-runner).

Note that the resulting libMLIR.so depends on LLVM, but
does not contain any LLVM components.  As a result, it
is necessary to link with libLLVM.so to avoid linkage
errors. So, libMLIR.so requires LLVM_BUILD_LLVM_DYLIB=on

FYI, Currently it appears that LLVM_LINK_LLVM_DYLIB is broken
because mlir-tblgen is linked against libLLVM.so and
and independent LLVM components

(updated by Stephen Neuendorffer)

Differential Revision: https://reviews.llvm.org/D73130

4 years ago[MLIR] Move from add_dependencies() to DEPENDS
Stephen Neuendorffer [Thu, 20 Feb 2020 18:48:51 +0000 (10:48 -0800)]
[MLIR] Move from add_dependencies() to DEPENDS

add_llvm_library and add_llvm_executable may need to create new targets with
appropriate dependencies.  As a result, it is not sufficient in some
configurations (namely LLVM_BUILD_LLVM_DYLIB=on) to only call
add_dependencies().  Instead, the explicit TableGen dependencies must
be passed to add_llvm_library() or add_llvm_executable() using the DEPENDS
keyword.

Differential Revision: https://reviews.llvm.org/D74930

4 years ago[MLIR] Remove redundant library dependencies
Stephen Neuendorffer [Thu, 20 Feb 2020 06:56:38 +0000 (22:56 -0800)]
[MLIR] Remove redundant library dependencies

In cmake, it is redundant to have a target list under target_link_libraries()
and add_dependency().  This patch removes the redundant dependency from
add_dependency().

Differential Revision: https://reviews.llvm.org/D74929

4 years ago[MLIR] Move from using target_link_libraries to LINK_LIBS for llvm libraries.
Stephen Neuendorffer [Tue, 18 Feb 2020 23:07:35 +0000 (15:07 -0800)]
[MLIR] Move from using target_link_libraries to LINK_LIBS for llvm libraries.

When compiling libLLVM.so, add_llvm_library() manipulates the link libraries
being used.  This means that when using add_llvm_library(), we need to pass
the list of libraries to be linked (using the LINK_LIBS keyword) instead of
using the standard target_link_libraries call.  This is preparation for
properly dealing with creating libMLIR.so as well.

Differential Revision: https://reviews.llvm.org/D74864

4 years ago[MLIR] Refactor library initialization of JitRunner.
Stephen Neuendorffer [Thu, 27 Feb 2020 17:47:37 +0000 (09:47 -0800)]
[MLIR] Refactor library initialization of JitRunner.

Previously, lib/Support/JitRunner.cpp was essentially a complete application,
performing all library initialization, along with dealing with command line
arguments and actually running passes.  This differs significantly from
mlir-opt and required a dependency on InitAllDialects.h.  This dependency
is significant, since it requires a dependency on all of the resulting
libraries.

This patch refactors the code so that tools are responsible for library
initialization, including registering all dialects, prior to calling
JitRunnerMain.  This places the concern about what dialect to support
with the end application, enabling more extensibility at the cost of
a small amount of code duplication between tools.  It also fixes
BUILD_SHARED_LIBS=on.

Differential Revision: https://reviews.llvm.org/D75272

4 years ago[MLIR] Refactor library handling for conversions.
Stephen Neuendorffer [Thu, 27 Feb 2020 00:31:14 +0000 (16:31 -0800)]
[MLIR] Refactor library handling for conversions.

Collect a list of conversion libraries in cmake, so we don't have to
list these explicitly in most binaries.

Differential Revision: https://reviews.llvm.org/D75222

4 years ago[MLIR] Refactor handling of dialect libraries
Stephen Neuendorffer [Wed, 26 Feb 2020 19:50:03 +0000 (11:50 -0800)]
[MLIR] Refactor handling of dialect libraries

Instead of creating extra libraries we don't really need, collect a
list of all dialects and use that instead.

Differential Revision: https://reviews.llvm.org/D75221

4 years ago[mlir] Fix typo
Jacques Pienaar [Fri, 28 Feb 2020 18:59:34 +0000 (10:59 -0800)]
[mlir] Fix typo

4 years agoAdd a pass that specializes parallel loops for easier unrolling and vectorization
Benjamin Kramer [Thu, 27 Feb 2020 12:24:27 +0000 (13:24 +0100)]
Add a pass that specializes parallel loops for easier unrolling and vectorization

This matches loops with a affine.min upper bound, limiting the trip
count to a constant, and rewrites them into two loops, one with constant
upper bound and one with variable upper bound. The assumption is that
the constant upper bound loop will be unrolled and vectorized, which is
preferable if this is the hot path.

Differential Revision: https://reviews.llvm.org/D75240

4 years ago[AST Matchers] Fix bug in 'optionally' matcher wherein all previous bindings are...
Yitzhak Mandelbaum [Fri, 28 Feb 2020 17:45:19 +0000 (12:45 -0500)]
[AST Matchers] Fix bug in 'optionally' matcher wherein all previous bindings are cleared when all inner matchers fail.

Summary: The implementation of 'optionally' doesn't preserve bindings when none of the submatchers succeed. This patch adds a regression test for that behavior and fixes it.

Reviewers: aaron.ballman, sbenza

Subscribers: cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D75365

4 years ago[DAGCombine] Fix alias analysis for unaligned accesses
David Green [Fri, 28 Feb 2020 18:31:45 +0000 (18:31 +0000)]
[DAGCombine] Fix alias analysis for unaligned accesses

The alias analysis in DAG Combine looks at the BaseAlign, the Offset and
the Size of two accesses, and determines if they are known to access
different parts of memory by the fact that they are different offsets
from inside that "alignment window". It does not seem to account for
accesses that are not a multiple of the size, and may overflow from one
alignment window into another.

For example in the test case we have a 19byte memset that is splits into
a 16 byte neon store and an unaligned 4 byte store with a 15 byte
offset. This 15byte offset (with a base align of 8) wraps around to the
next alignment windows. When compared to an access that is a 16byte
offset (of the same 4byte size and 8byte basealign), the two accesses
are said not to alias.

I've fixed this here by just ensuring that the offsets are a multiple of
the size, ensuring that they don't overlap by wrapping. Fixes PR45035,
which was exposed by the UseAA changes in the arm backend.

Differential Revision: https://reviews.llvm.org/D75238

4 years ago[VectorCombine] Fix assert on compare extract index
Austin Kerbow [Fri, 28 Feb 2020 18:08:59 +0000 (10:08 -0800)]
[VectorCombine] Fix assert on compare extract index

Extract index could be a differnet integral type.

Differential Revision: https://reviews.llvm.org/D75327

4 years ago[libc++] update GCC cherry-pick to build 4.8.5
Eric Fiselier [Fri, 28 Feb 2020 18:35:47 +0000 (13:35 -0500)]
[libc++] update GCC cherry-pick to build 4.8.5

4 years ago[SLP][NFC] Assert that tree entry operands completed when scheduler looks for depende...
Valery N Dmitriev [Thu, 27 Feb 2020 20:50:08 +0000 (12:50 -0800)]
[SLP][NFC] Assert that tree entry operands completed when scheduler looks for dependencies.

This change adds an assertion to prevent tricky bug related to recursive
approach of building vectorization tree. For loop below takes number of
operands directly from tree entry rather than from scalars.
If the entry at this moment turns out incomplete (i.e. not all operands set)
then not all the dependencies will be seen by the scheduler.
This can lead to failed scheduling (and thus failed vectorization)
for perfectly vectorizable tree.
Here is code example which is likely to fire the assertion:
for (i : VL0->getNumOperands()) {
  ...
  TE->setOperand(i, Operands);
  buildTree_rec(Operands, Depth + 1,...);
}

Correct way is two steps process: first set all operands to a tree entry
and then recursively process each operand.

Differential Revision: https://reviews.llvm.org/D75296

4 years ago[SLP]Update test checks, NFC.
Alexey Bataev [Fri, 28 Feb 2020 18:18:32 +0000 (13:18 -0500)]
[SLP]Update test checks, NFC.

4 years ago[X86] Recognize CVTPH2PS from STRICT_FP_EXTEND
Craig Topper [Fri, 28 Feb 2020 17:58:42 +0000 (09:58 -0800)]
[X86] Recognize CVTPH2PS from STRICT_FP_EXTEND

This should avoid scalarizing the cvtph2ps intrinsics with D75162

Differential Revision: https://reviews.llvm.org/D75304

4 years ago[lld][WebAssembly] Handle mixed strong and weak undefined symbols
Sam Clegg [Fri, 28 Feb 2020 01:32:22 +0000 (17:32 -0800)]
[lld][WebAssembly] Handle mixed strong and weak undefined symbols

When there are both strong and weak references to an undefined
symbol ensure that the strong reference prevails in the output symbol
generating the correct error.

Test case copied from lld/test/ELF/weak-and-strong-undef.s

Differential Revision: https://reviews.llvm.org/D75322

4 years agoDevirtualize a call on alloca without waiting for post inline cleanup and next Devirt...
Hiroshi Yamauchi [Fri, 28 Feb 2020 17:43:32 +0000 (09:43 -0800)]
Devirtualize a call on alloca without waiting for post inline cleanup and next DevirtSCCRepeatedPass iteration.

This aims to fix a missed inlining case.

If there's a virtual call in the callee on an alloca (stack allocated object) in
the caller, and the callee is inlined into the caller, the post-inline cleanup
would devirtualize the virtual call, but if the next iteration of
DevirtSCCRepeatedPass doesn't happen (under the new pass manager), which is
based on a heuristic to determine whether to reiterate, we may miss inlining the
devirtualized call.

This enables inlining in clang/test/CodeGenCXX/member-function-pointer-calls.cpp.

This is a second commit after a revert
https://reviews.llvm.org/rG4569b3a86f8a4b1b8ad28fe2321f936f9d7ffd43 and a fix
https://reviews.llvm.org/rG41e06ae7ba91.

Differential Revision: https://reviews.llvm.org/D69591

4 years ago[CallPromotionUtils] Add missing promotion legality check to tryPromoteCall.
Hiroshi Yamauchi [Thu, 27 Feb 2020 21:59:59 +0000 (13:59 -0800)]
[CallPromotionUtils] Add missing promotion legality check to tryPromoteCall.

Summary: This fixes the crash that led to the revert of D69591.

Reviewers: davidxl

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75307

4 years ago[SLP][NFC] Delete some unreachable code.
Valery N Dmitriev [Thu, 27 Feb 2020 18:57:06 +0000 (10:57 -0800)]
[SLP][NFC] Delete some unreachable code.

This patch deletes some dead code out of SLP vectorizer.
Couple of changes taken out of D57059 to slightly lighten it
plus one more similar case fixed.

Differential Revision: https://reviews.llvm.org/D75276

4 years agoRevert "[NFC][ARM] Update test"
Christopher Tetreault [Fri, 28 Feb 2020 16:36:27 +0000 (08:36 -0800)]
Revert "[NFC][ARM] Update test"

Summary:
There exists no corresponding code change for this commit, and this
commit causes downstream breakages.

This reverts commit 2db5547c016dbbd6acac3f3175937324f0095226.

Reviewers: samparker

Subscribers: kristof.beyls, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75358

4 years ago[mlir] [VectorOps] Add vector.broadcast to EDSC
aartbik [Fri, 28 Feb 2020 00:43:49 +0000 (16:43 -0800)]
[mlir] [VectorOps] Add vector.broadcast to EDSC

Reviewers: nicolasvasilache, andydavis1

Reviewed By: nicolasvasilache

Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75320

4 years ago[AVX512] Add strict-fp cvtph2ps constrained tests
Simon Pilgrim [Fri, 28 Feb 2020 16:54:45 +0000 (16:54 +0000)]
[AVX512] Add strict-fp cvtph2ps constrained tests

As suggested on D75162

4 years ago[F16C] Add strict-fp constrained tests
Simon Pilgrim [Fri, 28 Feb 2020 16:18:32 +0000 (16:18 +0000)]
[F16C] Add strict-fp constrained tests

As suggested on D75162

4 years ago[mlir] Add reifyReturnShape to shaped type OpInterface
Jacques Pienaar [Fri, 28 Feb 2020 16:37:09 +0000 (08:37 -0800)]
[mlir] Add reifyReturnShape to shaped type OpInterface

This call results in inserting operations that compute the return shape
dynamically for the operation.

4 years ago[Inliner] Inlining should honor nobuiltin attributes
Teresa Johnson [Thu, 6 Feb 2020 21:28:41 +0000 (13:28 -0800)]
[Inliner] Inlining should honor nobuiltin attributes

Summary:
Final patch in series to fix inlining between functions with different
nobuiltin attributes/options, which was specifically an issue in LTO.
See discussion on D61634 for background.

The prior patch in this series (D67923) enabled per-Function TLI
construction that identified the nobuiltin attributes.

Here I have allowed inlining to proceed if the callee's nobuiltins are a
subset of the caller's nobuiltins, but not in the reverse case, which
should be conservatively correct. This is controlled by a new option,
-inline-caller-superset-nobuiltin, which is enabled by default.

Reviewers: hfinkel, gchatelet, chandlerc, davidxl

Subscribers: arsenm, jvesely, nhaehnle, mehdi_amini, eraman, hiraditya, haicheng, dexonsmith, kerbowa, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D74162

4 years agoFix MSVC "32-bit shift implicitly converted to 64 bits" warning. NFCI.
Simon Pilgrim [Fri, 28 Feb 2020 15:19:57 +0000 (15:19 +0000)]
Fix MSVC "32-bit shift implicitly converted to 64 bits" warning. NFCI.

4 years ago[TargetLowering] SimplifyDemandedBits - fix SCALAR_TO_VECTOR knownbits bug
Simon Pilgrim [Fri, 28 Feb 2020 15:18:08 +0000 (15:18 +0000)]
[TargetLowering] SimplifyDemandedBits - fix SCALAR_TO_VECTOR knownbits bug

We can only report the knownbits for a SCALAR_TO_VECTOR node if we only demand the 0'th element - the upper elements are undefined and shouldn't be trusted.

This is causing a number of regressions that need addressing but we need to get the bugfix in first.

4 years ago[Transform][MemCpyOpt] Add missing DebugLoc to %tmpbitcast
Pierre-vh [Wed, 26 Feb 2020 13:41:33 +0000 (13:41 +0000)]
[Transform][MemCpyOpt] Add missing DebugLoc to %tmpbitcast

Fix for https://bugs.llvm.org/show_bug.cgi?id=37967

Differential Revision: https://reviews.llvm.org/D75173

4 years agoReland 7691790dfd1011d08f5468f63952d7690755aad4 with a MSAN fix
Krzysztof Parzyszek [Wed, 26 Feb 2020 15:56:12 +0000 (09:56 -0600)]
Reland 7691790dfd1011d08f5468f63952d7690755aad4 with a MSAN fix

In some cases when HexagonTargetLowering::allowsMemoryAccess returned
true, it did not set the "Fast" argument, leaving it uninitialized.

[Hexagon] Improve casting of boolean HVX vectors to scalars

- Mark memory access for bool vectors as disallowed in target lowering.
  This will prevent combining bitcasts of bool vectors with stores.
- Replace the actual bitcasting code with a faster version.
- Handle casting of v16i1 to i16.

4 years ago[ARM] MVE VMLAS
David Green [Fri, 28 Feb 2020 11:02:20 +0000 (11:02 +0000)]
[ARM] MVE VMLAS

This addes extra patterns for the VMLAS MVE instruction, which performs
Qda = Qda * Qn + Rm, a similar pattern to the existing VMLA. The sinking
of splat(Rm) into the loop is already performed, meaning we just need
extra Pat's in tablegen.

Differential Revision: https://reviews.llvm.org/D75115

4 years ago[ARM] Additional MVE VMLA tests. NFC
David Green [Thu, 27 Feb 2020 18:46:39 +0000 (18:46 +0000)]
[ARM] Additional MVE VMLA tests. NFC

4 years agoSkip TemplateSpecializedType in modernize-pass-by-value.
Karasev Nikita [Fri, 28 Feb 2020 14:17:16 +0000 (09:17 -0500)]
Skip TemplateSpecializedType in modernize-pass-by-value.

Existing 'modernize-pass-by-value' check works only with non template values in
initializers. Fixes PR37210.

4 years ago[cmake][msvc] Don't disable C4345 any more.
Simon Pilgrim [Fri, 28 Feb 2020 13:56:48 +0000 (13:56 +0000)]
[cmake][msvc] Don't disable C4345 any more.

This shouldn't be relevant now that we just support VS2017+.

4 years ago[Utils] Make some scripts directly executable
Jay Foad [Fri, 28 Feb 2020 13:39:54 +0000 (13:39 +0000)]
[Utils] Make some scripts directly executable

4 years ago[AMDGPU] Mark the scheduling model as complete
Jay Foad [Fri, 28 Feb 2020 13:22:44 +0000 (13:22 +0000)]
[AMDGPU] Mark the scheduling model as complete

4 years ago[AMDGPU] Update a comment missed in 74e2974ac6a
Jay Foad [Fri, 28 Feb 2020 13:20:55 +0000 (13:20 +0000)]
[AMDGPU] Update a comment missed in 74e2974ac6a

4 years agoFix buildbots after c074f5234d29439116f0e0be6033ea9331e85394.
Alexey Lapshin [Fri, 28 Feb 2020 12:06:53 +0000 (15:06 +0300)]
Fix buildbots after c074f5234d29439116f0e0be6033ea9331e85394.

Removed unused function getSectionByName() from dsymutil/DwarfStreamer.cpp.

4 years ago[clang-tidy] Added virtual isLanguageVersionSupported to ClangTidyCheck
Nathan James [Fri, 28 Feb 2020 13:03:30 +0000 (13:03 +0000)]
[clang-tidy] Added virtual isLanguageVersionSupported to ClangTidyCheck

Summary:
Motivated by [[ https://bugs.llvm.org/show_bug.cgi?id=45045 | Tune inspections to a specific C++ standard. ]]
Moves the isLanguageVersionSupported virtual function from `MakeSmartPtrCheck` to the base `ClangTidyCheck` class.
This will disable registering matchers or pp callbacks on unsupported language versions for a check.
Having it as a standalone function is cleaner than manually disabling the check in the register function and should hopefully
encourage check developers to actually restrict the check based on language version.
As an added bonus this could enable automatic detection of what language version a check runs on for the purpose of documentation generation

Reviewers: aaron.ballman, gribozavr2, Eugene.Zelenko, JonasToth, alexfh, hokein

Reviewed By: gribozavr2

Subscribers: xazax.hun, jkorous, arphaman, kadircet, usaxena95, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D75289

4 years ago[clang-format] Improve C# handling of spaces in square brackets
Jonathan Coe [Fri, 28 Feb 2020 12:44:15 +0000 (12:44 +0000)]
[clang-format] Improve C# handling of spaces in square brackets

Reviewers: MyDeveloperDay, krasimir

Reviewed By: krasimir

Subscribers: cfe-commits

Tags: #clang-format, #clang

Differential Revision: https://reviews.llvm.org/D75336

4 years ago[RISCV] Compress instructions based on function features
Simon Cook [Fri, 28 Feb 2020 11:52:55 +0000 (11:52 +0000)]
[RISCV] Compress instructions based on function features

When running under LTO, it is common to not specify the architecture
spec, which is used for setting up the target machine, and instead rely
on features specified in each function to generate the correct
instructions.

This works for the code generator, but the RISC-V backend uses the
AsmPrinter to do instruction compression, which does not see these
features but instead uses a MCSubtargetInfo object to see whether
compression is enabled. Since this is configured based on the
TargetMachine at startup, it will result in compressed instructions not
being emitted when it has not been given the 'c' TargetFeature, but the
function has it.

This changes the RISCVAsmPrinter to re-initialize the STI feature set
based on the current MachineFunction, such that compressed instructions
are now correctly emitted regardless of the method used to enable them.

Differential revision: https://reviews.llvm.org/D73339

4 years ago[gn build] Port 6af859dcca2
LLVM GN Syncbot [Fri, 28 Feb 2020 11:49:23 +0000 (11:49 +0000)]
[gn build] Port 6af859dcca2

4 years ago[ELF][LLD][ARM] Add missing REQUIRES: arm to tests
Peter Smith [Fri, 28 Feb 2020 11:44:32 +0000 (11:44 +0000)]
[ELF][LLD][ARM] Add missing REQUIRES: arm to tests

Fix buildbots that don't build ARM backend.

4 years ago[DebugInfo] Re-implement LexicalScopes dominance method, add unit tests
Jeremy Morse [Fri, 28 Feb 2020 10:41:23 +0000 (10:41 +0000)]
[DebugInfo] Re-implement LexicalScopes dominance method, add unit tests

Way back in D24994, the combination of LexicalScopes::dominates and
LiveDebugValues was identified as having worst-case quadratic complexity,
but it wasn't triggered by any code path at the time. I've since run into a
scenario where this occurs, in a very large basic block where large numbers
of inlined DBG_VALUEs are present.

The quadratic-ness comes from LiveDebugValues::join calling "dominates" on
every variable location, and LexicalScopes::dominates potentially touching
every instruction in a block to test for the presence of a scope. We have,
however, already computed the presence of scopes in blocks, in the
"InstrRanges" of each scope. This patch switches the dominates method to
examine whether a block is present in a scope's InsnRanges, avoiding
walking through the whole block.

At the same time, fix getMachineBasicBlocks to account for the fact that
InsnRanges can cover multiple blocks, and add some unit tests, as Lexical
Scopes didn't have any.

Differential revision: https://reviews.llvm.org/D73725

4 years agoLet EarlyCSE fold equivalent freeze instructions
Juneyoung Lee [Fri, 28 Feb 2020 09:47:57 +0000 (18:47 +0900)]
Let EarlyCSE fold equivalent freeze instructions

Summary:
This patch makes EarlyCSE fold equivalent freeze instructions.

Another optimization that I think will be useful is to remove freeze if its operand is used as a branch condition or at llvm.assume:

```
  %c = ...
  br i1 %c, label %A, ..
A:
  %d = freeze %c ; %d can be optimized to %c because %c cannot be poison or undef (or 'br %c' would be UB otherwise)
```

If it make sense for EarlyCSE to support this as well, I will make a patch for this.

Reviewers: spatel, reames, lebedev.ri

Reviewed By: lebedev.ri

Subscribers: lebedev.ri, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75334

4 years ago[LLD][ELF][ARM] Implement Thumb pc-relative relocations for adr and ldr
Peter Smith [Sun, 23 Feb 2020 19:46:46 +0000 (19:46 +0000)]
[LLD][ELF][ARM] Implement Thumb pc-relative relocations for adr and ldr

MC will now output the R_ARM_THM_PC8, R_ARM_THM_PC12 and
R_ARM_THM_PREL_11_0 relocations. These are short-ranged relocations that
are used to implement the adr rd, literal and ldr rd, literal pseudo
instructions.

The instructions use a new RelExpr called R_ARM_PCA in order to calculate
the required S + A - Pa expression, where Pa is AlignDown(P, 4) as the
instructions add their immediate to AlignDown(PC, 4). We also do not want
these relocations to generate or resolve against a PLT entry as the range
of these relocations is so short they would never reach.

The R_ARM_THM_PC8 has a special encoding convention for the relocation
addend, the immediate field is unsigned, yet the addend must be -4 to
account for the Thumb PC bias. The ABI (not the architecture) uses the
convention that the 8-byte immediate of 0xff represents -4.

Differential Revision: https://reviews.llvm.org/D75042

4 years ago[MC][ELF][ARM] Add relocations for some pc-relative fixups
Peter Smith [Sun, 23 Feb 2020 10:49:51 +0000 (10:49 +0000)]
[MC][ELF][ARM] Add relocations for some pc-relative fixups

Add ELF relocations for the following fixups:
fixup_thumb_adr_pcrel_10 -> R_ARM_THM_PC8
fixup_thumb_cp -> R_ARM_THM_PC8
fixup_t2_adr_pcrel_12 -> R_ARM_THM_PREL_11_0
fixup_t2_ldst_pcrel_12 -> R_ARM_THM_PC12

While these relocations are short-ranged there is support in the open
source ELF linker's in binutils and soon to be in LLD. MC will no longer
resolve pc-relative fixups to global symbols due to interpositioning
concerns. We can handle these at link time by implementing the relocations.

The R_ARM_THM_PC8 has some extra encoding rules for addends that llvm-mc
sidesteps by not supporting addends for these instructions, using the wide
Thumb 2 instruction if it is available. I think that this is a reasonable
compromise given that these are rare.

This partiall reverts D72892, the Thumb fixups no longer need to be
evaluated at assembly time.

Differential Revision: https://reviews.llvm.org/D75039

4 years ago[NFC][ARM] Add tests
Sam Parker [Fri, 28 Feb 2020 11:22:09 +0000 (11:22 +0000)]
[NFC][ARM] Add tests

4 years ago[AMDGPU] Precommit some scheduler related test updates
Jay Foad [Thu, 27 Feb 2020 21:16:46 +0000 (21:16 +0000)]
[AMDGPU] Precommit some scheduler related test updates

Summary:
The point of this is to make some tests with manual checks robust
against scheduler tweaks, so that only autogenerated test updates will
be required when pushing D68338 "[AMDGPU] Remove dubious logic in
bidirectional list scheduler".

Reviewers: arsenm, rampitec, vpykhtin

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, kerbowa, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D75302

4 years ago[RDA] Track implicit-defs
Sam Parker [Fri, 28 Feb 2020 11:14:42 +0000 (11:14 +0000)]
[RDA] Track implicit-defs

Ensure that we're recording implicit defs, as well as visiting implicit
uses and implicit defs when we're walking through operands.

Differential Revision: https://reviews.llvm.org/D75185