platform/upstream/llvm.git
3 years ago[flang][driver] Report prescanning diags during syntax-only parsing
Faris Rehman [Wed, 27 Jan 2021 09:45:56 +0000 (09:45 +0000)]
[flang][driver] Report prescanning diags during syntax-only parsing

Ensure diagnostics from the prescanner are reported when running `flang-new -fsyntax-only` (i.e. only syntax parsing).
This keeps the diagnostics output of flang-new consistent with `f18 -fparse-only` when running the syntax parsing action, ParseSyntaxOnlyAction.

Summary of changes:
- Modify ParseSyntaxOnlyAction::ExecuteAction to report diagnostics

Differential Revision: https://reviews.llvm.org/D95220

3 years agoFix "not all control paths return a value" warning. NFCI.
Simon Pilgrim [Wed, 27 Jan 2021 10:14:54 +0000 (10:14 +0000)]
Fix "not all control paths return a value" warning. NFCI.

3 years ago[AMDGPU] Write "GFX6-GFX9" instead of "GFX6-9" in docs
Jay Foad [Tue, 26 Jan 2021 17:45:41 +0000 (17:45 +0000)]
[AMDGPU] Write "GFX6-GFX9" instead of "GFX6-9" in docs

... and similarly for some other cases. This is for consistency and to
make it easier to search for mentions of a particular architecture.

Differential Revision: https://reviews.llvm.org/D95453

3 years ago[ARM] Add neon FP16 scalar_to_vector patterns.
David Green [Wed, 27 Jan 2021 09:59:15 +0000 (09:59 +0000)]
[ARM] Add neon FP16 scalar_to_vector patterns.

This adds some simple fp16 scalar_to_vector patterns, preventing a
selection failure if this came up.

Differential Revision: https://reviews.llvm.org/D95427

3 years ago[Test][AArch64] Use named vregs in overflow legalization tests. NFC
Cassie Jones [Wed, 27 Jan 2021 08:04:18 +0000 (03:04 -0500)]
[Test][AArch64] Use named vregs in overflow legalization tests. NFC

3 years ago[AArch64][GlobalISel] Make G_SADDE and G_SSUBE legal
Cassie Jones [Wed, 27 Jan 2021 09:36:12 +0000 (04:36 -0500)]
[AArch64][GlobalISel] Make G_SADDE and G_SSUBE legal

This makes G_SADDE and G_SSUBE legal in preparation for further work
legalizing overflowing operations. It's fine that they don't have an
instruction selector implementation yet, because G_UADDE and G_USUBE are
already legal on AArch64 without an instruction selector implementation. This
completes the set of G_[SU]{ADD,SUB}[EO] operations on AArch64.

Reviewed By: paquette

Differential Revision: https://reviews.llvm.org/D95325

3 years ago[clang] Fix signedness in vector bitcast evaluation
Sven van Haastregt [Wed, 27 Jan 2021 09:30:26 +0000 (09:30 +0000)]
[clang] Fix signedness in vector bitcast evaluation

The included test case triggered a sign assertion on the result in
`Success()`.  This was caused by the APSInt created for a bitcast
having its signedness bit inverted.  The second APSInt constructor
argument is `isUnsigned`, so invert the result of
`isSignedIntegerType`.

Relanding this patch after reverting.  The test case had to be updated
to be insensitive to 32/64-bit extractelement indices.

Differential Revision: https://reviews.llvm.org/D95135

3 years ago[libc][NFC] Use a end of list marker for cpu feature detection.
Siva Chandra Reddy [Wed, 27 Jan 2021 09:18:58 +0000 (01:18 -0800)]
[libc][NFC] Use a end of list marker for cpu feature detection.

Without this, the array can end up being an empty array leading to
compiler failures.

3 years ago[OpenMP] libomp: fix build by clang-cl with vs2019
AndreyChurbanov [Wed, 27 Jan 2021 09:18:15 +0000 (12:18 +0300)]
[OpenMP] libomp: fix build by clang-cl with vs2019

Problem reported by Joseph Shen <joseph.smeng@gmail.com>.
The patch changes *(&<atomic-var>) to (&<atomic-var>)->load().

Differential Revision: https://reviews.llvm.org/D95485

3 years ago[clang][cli] Port LangOpts to marshalling system, pt.2
Jan Svoboda [Wed, 27 Jan 2021 08:00:16 +0000 (09:00 +0100)]
[clang][cli] Port LangOpts to marshalling system, pt.2

Port some miscellaneous language options to the marshalling system for oautomatic command line parsing and generation.

Reviewed By: dexonsmith

Differential Revision: https://reviews.llvm.org/D95347

3 years ago[mlir] Extend semantic of OffsetSizeAndStrideOpInterface.
Nicolas Vasilache [Wed, 27 Jan 2021 08:06:40 +0000 (08:06 +0000)]
[mlir] Extend semantic of OffsetSizeAndStrideOpInterface.

OffsetSizeAndStrideOpInterface now have the ability to specify only a leading subset of
offset, sizes, strides operands/attributes.
The size of that leading subset must be limited by the corresponding entry in `getArrayAttrMaxRanks` to avoid overflows.
Missing trailing dimensions are assumed to span the whole range (i.e. [0 .. dim)).
This brings more natural semantics to slice-like op on top of subview and is a simplifies to removing all uses of SliceOp in dependent projects.

Differential revision: https://reviews.llvm.org/D95441

3 years agoFix an error about implicit fallthrough during self build - new tag for ittapi.
Alexey Bader [Wed, 27 Jan 2021 05:53:38 +0000 (08:53 +0300)]
Fix an error about implicit fallthrough during self build - new tag for ittapi.

A fix has been implemented in the ittap repo to fix an error about implicit fallthrough in a switch that was occurring during self build.
A new tag has been created for that fix. This is to update the tag.

Reviewed By: bader

Differential Revision: https://reviews.llvm.org/D95462

Patch by Zahira Ammarguellat.

3 years ago[clang-format] Avoid considering include directive as a template closer.
Marek Kurdej [Wed, 27 Jan 2021 08:14:22 +0000 (09:14 +0100)]
[clang-format] Avoid considering include directive as a template closer.

This fixes a bug [[ http://llvm.org/PR48891 | PR48891 ]] introduced in D93839 where:
```
#include <stdint.h>
namespace rep {}
```
got formatted as
```
#include <stdint.h>
namespace rep {
}
```

Reviewed By: MyDeveloperDay, leonardchan

Differential Revision: https://reviews.llvm.org/D95479

3 years ago[clang][cli] Port LangOpts to marshalling system, pt.1
Jan Svoboda [Wed, 27 Jan 2021 07:47:54 +0000 (08:47 +0100)]
[clang][cli] Port LangOpts to marshalling system, pt.1

Port some miscellaneous language options to the marshalling system for oautomatic command line parsing and generation.

Reviewed By: dexonsmith

Differential Revision: https://reviews.llvm.org/D95346

3 years ago[mlir][Linalg] Add canonicalization for init_tensor -> subtensor op.
MaheshRavishankar [Wed, 27 Jan 2021 07:21:33 +0000 (23:21 -0800)]
[mlir][Linalg] Add canonicalization for init_tensor -> subtensor op.

Differential Revision: https://reviews.llvm.org/D95305

3 years ago[llvm-objdump] Use append_range (NFC)
Kazu Hirata [Wed, 27 Jan 2021 04:00:19 +0000 (20:00 -0800)]
[llvm-objdump] Use append_range (NFC)

3 years ago[MemorySSA] Use ListSeparator (NFC)
Kazu Hirata [Wed, 27 Jan 2021 04:00:17 +0000 (20:00 -0800)]
[MemorySSA] Use ListSeparator (NFC)

3 years ago[AMDGPU] Forward-declare TargetRegisterClass (NFC)
Kazu Hirata [Wed, 27 Jan 2021 04:00:16 +0000 (20:00 -0800)]
[AMDGPU] Forward-declare TargetRegisterClass (NFC)

AMDGPUInstructionSelector.h needs TargetRegisterClass but relies on a
forward declaration of TargetRegisterClass in InstructionSelector.h.
This patch adds a forward declaration right in
AMDGPUInstructionSelector.h.

While we are at it, this patch removes the one in
InstructionSelector.h, where it is unnecessary.

3 years ago[TableGen] Add isContradictoryImpl implementation to CheckCondCodeMatcher and CheckCh...
Craig Topper [Wed, 27 Jan 2021 03:41:52 +0000 (19:41 -0800)]
[TableGen] Add isContradictoryImpl implementation to CheckCondCodeMatcher and CheckChild2CondCodeMatcher.

This enables better pattern factoring in the RISCV ISel table.

3 years agoBump the trunk major version to 13 llvmorg-13-init
Tom Stellard [Wed, 27 Jan 2021 03:37:08 +0000 (19:37 -0800)]
Bump the trunk major version to 13

and clear the release notes.

3 years agoFrontend: Use early returns in CompilerInstance::clearOutputFiles, NFC
Duncan P. N. Exon Smith [Wed, 27 Jan 2021 03:27:32 +0000 (19:27 -0800)]
Frontend: Use early returns in CompilerInstance::clearOutputFiles, NFC

Use early returns in `CompilerInstance::clearOutputFiles` to clarify the
logic, and rename `ec` to `EC` as a drive-by.

No functionality change.

3 years agoRename clang/test/Frontend/output-{failures,paths}.c, NFC
Duncan P. N. Exon Smith [Wed, 27 Jan 2021 03:26:24 +0000 (19:26 -0800)]
Rename clang/test/Frontend/output-{failures,paths}.c, NFC

A follow up patch will add a few success cases here; rename it to
`output-paths.c` instead of `output-failures.c`.

3 years ago[gn build] Port bb9eb1982980
LLVM GN Syncbot [Wed, 27 Jan 2021 01:23:23 +0000 (01:23 +0000)]
[gn build] Port bb9eb1982980

3 years ago[OpenMP][NVPTX] Drop dependence on CUDA to build NVPTX `deviceRTLs`
Shilei Tian [Wed, 27 Jan 2021 01:21:27 +0000 (20:21 -0500)]
[OpenMP][NVPTX] Drop dependence on CUDA to build NVPTX `deviceRTLs`

With D94745, we no longer use CUDA SDK to compile `deviceRTLs`. Therefore,
many CMake code in the project is useless. This patch cleans up unnecessary code
and also drops the requirement to build NVPTX `deviceRTLs`. CUDA detection is
still being used however to determine whether we need to involve the tests. Auto
detection of compute capability is enabled by default and can be disabled by
setting CMake variable `LIBOMPTARGET_NVPTX_AUTODETECT_COMPUTE_CAPABILITY=OFF`.
If auto detection is enabled, and CUDA is also valid, it will only build the
bitcode library for the detected version; otherwise, all variants supported will
be generated. One drawback of this patch is, we now generate 96 variants of
bitcode library, and totally 1485 files to be built with a clean build on a
non-CUDA system. `LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=""` can be used to
disable building NVPTX `deviceRTLs`.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D95466

3 years ago[RISCV] Add rv64 run lines to rv32 MC layer tests for B extension
Craig Topper [Wed, 27 Jan 2021 01:03:21 +0000 (17:03 -0800)]
[RISCV] Add rv64 run lines to rv32 MC layer tests for B extension

Remove common instructions from rv64 tests since they are now
covered by the rv64 run lines in the rv32 tests.

Add rv32-only* tests for a few cases that aren't common between
r32 and rv64.

Addresses review feedback from D95150.

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D95272

3 years agoSupport for instrumenting only selected files or functions
Petr Hosek [Fri, 15 Jan 2021 09:14:37 +0000 (01:14 -0800)]
Support for instrumenting only selected files or functions

This change implements support for applying profile instrumentation
only to selected files or functions. The implementation uses the
sanitizer special case list format to select which files and functions
to instrument, and relies on the new noprofile IR attribute to exclude
functions from instrumentation.

Differential Revision: https://reviews.llvm.org/D94820

3 years ago[libc++] Give `MoveOnly` all six comparison operators, not just == and <.
Arthur O'Dwyer [Mon, 25 Jan 2021 22:00:14 +0000 (17:00 -0500)]
[libc++] Give `MoveOnly` all six comparison operators, not just == and <.

Split out of D93512.

3 years ago[OpenMP] Modify OMP_ALLOCATOR environment variable
Nawrin Sultana [Mon, 2 Nov 2020 22:17:37 +0000 (16:17 -0600)]
[OpenMP] Modify OMP_ALLOCATOR environment variable

This patch sets the def-allocator-var ICV based on the environment variables
provided in OMP_ALLOCATOR. Previously, only allowed value for OMP_ALLOCATOR
was a predefined memory allocator. OpenMP 5.1 specification allows predefined
memory allocator, predefined mem space, or predefined mem space with traits in
OMP_ALLOCATOR. If an allocator can not be created using the provided environment
variables, the def-allocator-var is set to omp_default_mem_alloc.

Differential Revision: https://reviews.llvm.org/D94985

3 years ago[libomptarget][cuda] Handle missing _v2 symbols gracefully
Jon Chesterfield [Wed, 27 Jan 2021 00:22:28 +0000 (00:22 +0000)]
[libomptarget][cuda] Handle missing _v2 symbols gracefully

[libomptarget][cuda] Handle missing _v2 symbols gracefully

Follow on from D95367. Dlsym the _v2 symbols if present, otherwise use the
unsuffixed version. Builds a hashtable for the check, can revise for zero
heap allocations later if necessary.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D95415

3 years ago[gn build] fix get.py change
Nico Weber [Wed, 27 Jan 2021 00:20:23 +0000 (19:20 -0500)]
[gn build] fix get.py change

3 years ago[gn build] restore build command removed in 9595a7ff55b6 for platforms without prebuilts
Nico Weber [Wed, 27 Jan 2021 00:19:19 +0000 (19:19 -0500)]
[gn build] restore build command removed in 9595a7ff55b6 for platforms without prebuilts

3 years agoDisable rosegment for old Android versions.
Dan Albert [Wed, 27 Jan 2021 00:04:56 +0000 (16:04 -0800)]
Disable rosegment for old Android versions.

The unwinder used by the crash handler on versions of Android prior to
API 29 did not correctly handle binaries built with rosegment, which is
enabled by default for LLD. Android only supports LLD, so it's not an
issue that this flag is not accepted by other linkers.

Reviewed By: srhines

Differential Revision: https://reviews.llvm.org/D95166

3 years agollvm-lib: Pull error printing code out of two functions
Nico Weber [Wed, 27 Jan 2021 00:11:56 +0000 (19:11 -0500)]
llvm-lib: Pull error printing code out of two functions

Slightly changes the output in error code, but no behavior change in
normal use. This is for preparation for using these two functions
elsewhere.

3 years ago[libomptarget][NFC] Avoid gcc 5/6 issue with lambda captures.
Vyacheslav Zakharin [Tue, 26 Jan 2021 22:45:40 +0000 (14:45 -0800)]
[libomptarget][NFC] Avoid gcc 5/6 issue with lambda captures.

Differential Revision: https://reviews.llvm.org/D95486

3 years agoFrontend: Fix layering between create{,Default}OutputFile, NFC
Duncan P. N. Exon Smith [Sat, 21 Nov 2020 02:04:32 +0000 (18:04 -0800)]
Frontend: Fix layering between create{,Default}OutputFile, NFC

Fix layering between `CompilerInstance::createDefaultOutputFile` and the
two versions of `createOutputFile`.

- Add missing configuration flags to `createDefaultOutputFile` so that
  GeneratePCHAction and GenerateModuleFromModuleMapAction can use it.
  They previously promised that temporary files were turned on; now
  `createDefaultOutputFile` handles that logic.
- Lift the logic handling `InFile` and `Extension` to
  `createDefaultOutputFile`, since it's only the callers of that
  function that are using it.
- Rename the deeper of the two `createOutputFile`s to
  `createOutputFileImpl` and make it private to `CompilerInstance` (to
  prove that no one else is using it).
- Sink the logic for adding to `CompilerInstance::OutputFiles` down to
  `createOutputFileImpl`, allowing two "optional" (but always used)
  `std::string*` out parameters to be removed.
- Instead of passing a `std::error_code` out parameter into
  `createOutputFileImpl`, have it return `Expected<>`.
- As a drive-by, inline `CompilerInstance::addOutputFile` into its only
  caller, `createOutputFileImpl`.

Clean layering makes it easier for a future commit to extract
`createOutputFileImpl` out of `CompilerInstance`.

Differential Revision: https://reviews.llvm.org/D93248

3 years ago[llc] Add reportError helper and canonicalize error messages
Fangrui Song [Tue, 26 Jan 2021 23:33:37 +0000 (15:33 -0800)]
[llc] Add reportError helper and canonicalize error messages

3 years agoFrontend: Simplify handling of non-seeking streams in CompilerInstance, NFC
Duncan P. N. Exon Smith [Tue, 15 Dec 2020 01:46:11 +0000 (17:46 -0800)]
Frontend: Simplify handling of non-seeking streams in CompilerInstance, NFC

Add a new `raw_pwrite_ostream` variant, `buffer_unique_ostream`, which
is like `buffer_ostream` but with unique ownership of the stream it's
wrapping. Use this in CompilerInstance to simplify the ownership of
non-seeking output streams, avoiding logic sprawled around to deal with
them specially.

This also simplifies future work to encapsulate output files in a
different class.

Differential Revision: https://reviews.llvm.org/D93260

3 years ago[GlobalISel] Implement computeKnownBits for G_SEXT_INREG
Jessica Paquette [Tue, 26 Jan 2021 22:39:39 +0000 (14:39 -0800)]
[GlobalISel] Implement computeKnownBits for G_SEXT_INREG

Just use the existing `Known.sextInReg` implementation.

- Update KnownBitsTest.cpp.
- Update combine-redundant-and.mir for a more concrete example.

Differential Revision: https://reviews.llvm.org/D95484

3 years agoSalvage debug info for function arguments in coro-split funclets.
Adrian Prantl [Tue, 26 Jan 2021 22:30:10 +0000 (14:30 -0800)]
Salvage debug info for function arguments in coro-split funclets.

This patch improves the availability for variables stored in the
coroutine frame by emitting an alloca to hold the pointer to the frame
object and rewriting dbg.declare intrinsics to point inside the frame
object using salvaged DIExpressions. Finally, a new alloca is created
in the funclet to hold the FramePtr pointer to ensure that it is
available throughout the entire function at -O0.

This path also effectively reverts D90772. The testcase updates
highlight nicely how every removed CHECK for a dbg.value is preceded
by a new CHECK for a dbg.declare.

Thanks to JunMa, Yifeng, and Bruno for their thoughtful reviews!

Differential Revision: https://reviews.llvm.org/D93497

rdar://71866936

3 years agoFrontend: Fix memory leak in CompilerInstance::setVerboseOutputStream
Duncan P. N. Exon Smith [Sat, 21 Nov 2020 03:13:19 +0000 (19:13 -0800)]
Frontend: Fix memory leak in CompilerInstance::setVerboseOutputStream

Found this memory leak in `CompilerInstance::setVerboseOutputStream` by
inspection; it looks like this wasn't previously exercised, since it was
never called twice.

Differential Revision: https://reviews.llvm.org/D93249

3 years ago[ARM] Fix STRT/STRHT/STRBT input/output operands.
Zhuojia Shen [Tue, 26 Jan 2021 22:00:58 +0000 (14:00 -0800)]
[ARM] Fix STRT/STRHT/STRBT input/output operands.

STRT, STRHT, and STRBT are store instructions and their source register
$Rt should be treated as an input operand instead of an output operand.
This should fix things (e.g., liveness tracking in LivePhysRegs) if
these instructions were used in CodeGen.

Differential Revision: https://reviews.llvm.org/D95074

3 years ago[NewPM] Add ExtraVectorizerPasses support
Bjorn Pettersson [Fri, 22 Jan 2021 23:54:04 +0000 (00:54 +0100)]
[NewPM] Add ExtraVectorizerPasses support

As it looks like NewPM generally is using SimpleLoopUnswitch
instead of LoopUnswitch, this patch also use SimpleLoopUnswitch
in the ExtraVectorizerPasses sequence (compared with LegacyPM
which use the LoopUnswitch pass).

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D95457

3 years ago[libomptarget][NFC] Use portable printf format specifiers.
Vyacheslav Zakharin [Tue, 26 Jan 2021 21:52:53 +0000 (13:52 -0800)]
[libomptarget][NFC] Use portable printf format specifiers.

Differential Revision: https://reviews.llvm.org/D95476

3 years ago[InstCombine] Preserve FMF for powi simplifications.
Valery N Dmitriev [Tue, 26 Jan 2021 19:05:16 +0000 (11:05 -0800)]
[InstCombine] Preserve FMF for powi simplifications.

Differential Revision: https://reviews.llvm.org/D95455

3 years ago[NFC] Show instcombine powi simplifications drop FMF
Valery N Dmitriev [Tue, 26 Jan 2021 18:59:27 +0000 (10:59 -0800)]
[NFC] Show instcombine powi simplifications drop FMF

Differential Revision: https://reviews.llvm.org/D95454

3 years ago[X86] In shrinkAndImmediate, place the new constant into the topological sort.
Craig Topper [Tue, 26 Jan 2021 19:28:25 +0000 (11:28 -0800)]
[X86] In shrinkAndImmediate, place the new constant into the topological sort.

Revert the change to use APInt::isSignedIntN from
5ff5cf8e057782e3e648ecf5ccf1d9990b53ee90.

Its clear that the games we were playing to avoid the topological
sort aren't working. So just fix it once and for all.

Fixes PR48888.

3 years ago[NFC][lit] Cleanup code using string interpolation
Julian Lettner [Tue, 26 Jan 2021 18:54:16 +0000 (10:54 -0800)]
[NFC][lit] Cleanup code using string interpolation

LLVM now requires Python 3.6, so we can use string interpolation to make
code more readable.

3 years ago[GlobalISel][IRTranslator] Ignore the llvm.experimental.noalias.scope.decl intrinsic.
Amara Emerson [Tue, 26 Jan 2021 20:54:41 +0000 (12:54 -0800)]
[GlobalISel][IRTranslator] Ignore the llvm.experimental.noalias.scope.decl intrinsic.

These don't generate any code.

3 years ago[OpenMP][Libomptarget] Fix cmake error on remote plugin
Atmn Patel [Tue, 26 Jan 2021 20:56:48 +0000 (15:56 -0500)]
[OpenMP][Libomptarget] Fix cmake error on remote plugin

Requiring 3.15 causes a build breakage, I'm sure none of the contents actually require
3.15 or above.

Differential Revision: https://reviews.llvm.org/D95474

3 years ago[gn build] Port 1e634f3952aa
LLVM GN Syncbot [Tue, 26 Jan 2021 20:48:31 +0000 (20:48 +0000)]
[gn build] Port 1e634f3952aa

3 years ago[llvm-elfabi] Fix test after D95140
Fangrui Song [Tue, 26 Jan 2021 20:45:45 +0000 (12:45 -0800)]
[llvm-elfabi] Fix test after D95140

3 years ago[libomptarget][cuda] Gracefully handle missing cuda library
Jon Chesterfield [Tue, 26 Jan 2021 20:43:06 +0000 (20:43 +0000)]
[libomptarget][cuda] Gracefully handle missing cuda library

[libomptarget][cuda] Gracefully handle missing cuda library

If using dynamic cuda, and it failed to load, it is not safe to call
cuGetErrorString.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D95412

3 years ago[libomptarget][cuda] Only run tests when sure there is cuda available
Jon Chesterfield [Tue, 26 Jan 2021 20:41:06 +0000 (20:41 +0000)]
[libomptarget][cuda] Only run tests when sure there is cuda available

[libomptarget][cuda] Only run tests when sure there is cuda available

Prior to D95155, building the cuda plugin implied cuda was installed locally.
With that change, every machine can build a cuda plugin, but they won't all have
cuda and/or an nvptx card installed locally.

This change enables the nvptx tests when either:
- libcuda is present
- the user has forced use of the dlopen stub

The default case when there is no cuda detected will no longer attempt to
run the tests on nvptx hardware, as was the case before D95155.

Reviewed By: jdoerfert, ronlieb

Differential Revision: https://reviews.llvm.org/D95467

3 years ago[OpenMP][Libomptarget] Introduce Remote Offloading Plugin
Atmn Patel [Sat, 23 Jan 2021 17:23:50 +0000 (12:23 -0500)]
[OpenMP][Libomptarget] Introduce Remote Offloading Plugin

This introduces a remote offloading plugin for libomptarget. This
implementation relies on gRPC and protobuf, so this library will only
build if both libraries are available on the system. The corresponding
server is compiled to `openmp-offloading-server`.

This is a large change, but the only way to split this up is into RTL/server
but I fear that could introduce an inconsistency amongst them.

Ideally, tests for this should be added to the current ones that but that is
problematic for at least one reason. Given that libomptarget registers plugin
on a first-come-first-serve basis, if we wanted to offload onto a local x86
through a different process, then we'd have to either re-order the plugin list
in `rtl.cpp` (which is what I did locally for testing) or find a better
solution for runtime plugin registration in libomptarget.

Differential Revision: https://reviews.llvm.org/D95314

3 years ago[llvm-elfabi] Support ELF file that lacks .gnu.hash section
Haowei Wu [Tue, 26 Jan 2021 19:34:51 +0000 (11:34 -0800)]
[llvm-elfabi] Support ELF file that lacks .gnu.hash section

Before this change, when reading ELF file, elfabi determines number of
entries in .dynsym by reading the .gnu.hash section. This change makes
elfabi read section headers directly first. This change allows elfabi
works on ELF files which do not have .gnu.hash sections.

Differential Revision: https://reviews.llvm.org/D93362

3 years ago[libc++] Fix oss-fuzz build
Louis Dionne [Tue, 26 Jan 2021 20:30:42 +0000 (15:30 -0500)]
[libc++] Fix oss-fuzz build

3 years agoAdd -fbinutils-version= to gate ELF features on the specified binutils version
Fangrui Song [Tue, 26 Jan 2021 20:28:23 +0000 (12:28 -0800)]
Add -fbinutils-version= to gate ELF features on the specified binutils version

There are two use cases.

Assembler
We have accrued some code gated on MCAsmInfo::useIntegratedAssembler().  Some
features are supported by latest GNU as, but we have to use
MCAsmInfo::useIntegratedAs() because the newer versions have not been widely
adopted (e.g. SHF_LINK_ORDER 'o' and 'unique' linkage in 2.35, --compress-debug-sections= in 2.26).

Linker
We want to use features supported only by LLD or very new GNU ld, or don't want
to work around older GNU ld. We currently can't represent that "we don't care
about old GNU ld".  You can find such workarounds in a few other places, e.g.
Mips/MipsAsmprinter.cpp PowerPC/PPCTOCRegDeps.cpp X86/X86MCInstrLower.cpp
AArch64 TLS workaround for R_AARCH64_TLSLD_MOVW_DTPREL_* (PR ld/18276),
R_AARCH64_TLSLE_LDST8_TPREL_LO12 (https://bugs.llvm.org/show_bug.cgi?id=36727 https://sourceware.org/bugzilla/show_bug.cgi?id=22969)

Mixed SHF_LINK_ORDER and non-SHF_LINK_ORDER components (supported by LLD in D84001;
GNU ld feature request https://sourceware.org/bugzilla/show_bug.cgi?id=16833 may take a while before available).
This feature allows to garbage collect some unused sections (e.g. fragmented .gcc_except_table).

This patch adds `-fbinutils-version=` to clang and `-binutils-version` to llc.
It changes one codegen place in SHF_MERGE to demonstrate its usage.
`-fbinutils-version=2.35` means the produced object file does not care about GNU
ld<2.35 compatibility. When `-fno-integrated-as` is specified, the produced
assembly can be consumed by GNU as>=2.35, but older versions may not work.

`-fbinutils-version=none` means that we can use all ELF features, regardless of
GNU as/ld support.

Both clang and llc need `parseBinutilsVersion`. Such command line parsing is
usually implemented in `llvm/lib/CodeGen/CommandFlags.cpp` (LLVMCodeGen),
however, ClangCodeGen does not depend on LLVMCodeGen. So I add
`parseBinutilsVersion` to `llvm/lib/Target/TargetMachine.cpp` (LLVMTarget).

Differential Revision: https://reviews.llvm.org/D85474

3 years agoRevert "Support for instrumenting only selected files or functions"
Petr Hosek [Tue, 26 Jan 2021 20:25:28 +0000 (12:25 -0800)]
Revert "Support for instrumenting only selected files or functions"

This reverts commit 4edf35f11a9e20bd5df3cb47283715f0ff38b751 because
the test fails on Windows bots.

3 years agoMake SBDebugger::CreateTargetWithFileAndArch work with lldb::LLDB_DEFAULT_ARCH
Jim Ingham [Tue, 26 Jan 2021 20:15:09 +0000 (12:15 -0800)]
Make SBDebugger::CreateTargetWithFileAndArch work with lldb::LLDB_DEFAULT_ARCH

Second try, handling both a bogus arch string and the "null file & arch" used
to create an empty but valid target.
Also check in that case before logging (previously the logging would have
crashed.)

3 years ago[flang][openacc][NFC] Organize clause validity tests by directive
Valentin Clement [Tue, 26 Jan 2021 19:53:50 +0000 (14:53 -0500)]
[flang][openacc][NFC] Organize clause validity tests by directive

Split the tests from acc-clause-validity.f90 in dedicated files by directives.
The file acc-clause-validity.f90 was getting too big to be correctly maintained.
Tests are identical.

Reviewed By: SouraVX

Differential Revision: https://reviews.llvm.org/D95328

3 years agoCGDebugInfo CreatedLimitedType: Drop file/line for RecordType with invalid location
Fangrui Song [Tue, 26 Jan 2021 19:53:25 +0000 (11:53 -0800)]
CGDebugInfo CreatedLimitedType: Drop file/line for RecordType with invalid location

For Clang synthesized `__va_list_tag` (`CreateX86_64ABIBuiltinVaListDecl`),
its DW_AT_decl_file/DW_AT_decl_line are arbitrarily set from `CurLoc`.

In a stage 2 `-DCMAKE_BUILD_TYPE=Debug` clang build, I observe that
in driver.cpp, DW_AT_decl_file/DW_AT_decl_line may be set to an `#include` line
(the transitively included file uses va_arg (`__builtin_va_arg`)).
This seems arbitrary. Drop that.

Reviewed By: #debug-info, dblaikie

Differential Revision: https://reviews.llvm.org/D94735

3 years agoCGDebugInfo: Drop Loc.isInvalid() special case from getLineNumber
Fangrui Song [Tue, 26 Jan 2021 19:44:41 +0000 (11:44 -0800)]
CGDebugInfo: Drop Loc.isInvalid() special case from getLineNumber

`getLineNumber()` picks CurLoc if the parameter is invalid. This appears to
mainly work around missing SourceLocation information for some constructs, but
sometimes adds unintended locations.

* For `CodeGenObjC/debug-info-blocks.m`, `CurLoc` has been advanced to the closing brace. The debug line of `ImplicitVarParameter` is set to the line of `}` because this implicit parameter has an invalid `SourceLocation`. The debug line is a bit arbitrary - perhaps the location of `^{` is better.
* The file/line of Clang synthesized `__va_list_tag` is arbitrarily attached a `#include` line. D94735

Drop the special case to make getLineNumber less magic and add CurLoc fallback in its callers instead.

Tested with stage 2 -DCMAKE_BUILD_TYPE=Debug clang, byte identical.

Reviewed By: #debug-info, aprantl

Differential Revision: https://reviews.llvm.org/D94391

3 years ago[AMDGPU] Update subtarget features for new target ID support
Austin Kerbow [Mon, 30 Nov 2020 17:06:35 +0000 (09:06 -0800)]
[AMDGPU] Update subtarget features for new target ID support

Support for XNACK and SRAMECC is not static on some GPUs. We must be able
to differentiate between different scenarios for these dynamic subtarget
features.

The possible settings are:

- Unsupported: The GPU has no support for XNACK/SRAMECC.
- Any: Preference is unspecified. Use conservative settings that can run anywhere.
- Off: Request support for XNACK/SRAMECC Off
- On: Request support for XNACK/SRAMECC On

GCNSubtarget will track the four options based on the following criteria. If
the subtarget does not support XNACK/SRAMECC we say the setting is
"Unsupported". If no subtarget features for XNACK/SRAMECC are requested we
must support "Any" mode. If the subtarget features XNACK/SRAMECC exist in the
feature string when initializing the subtarget, the settings are "On/Off".

The defaults are updated to be conservatively correct, meaning if no setting
for XNACK or SRAMECC is explicitly requested, defaults will be used which
generate code that can be run anywhere. This corresponds to the "Any" setting.

Differential Revision: https://reviews.llvm.org/D85882

3 years ago[OpenMP][Libomptarget] Introduce changes to support remote plugin
Atmn [Tue, 26 Jan 2021 19:19:10 +0000 (14:19 -0500)]
[OpenMP][Libomptarget] Introduce changes to support remote plugin

In order to support remote execution, we need to be able to send the
target binary description to the remote host for registration (and
consequent deregistration). To support this, I added these two
optional new functions to the plugin API:
- `__tgt_rtl_register_lib`
- `__tgt_rtl_unregister_lib`

These functions will be called to properly manage the instance of
libomptarget running on the remote host.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D93293

3 years ago[gn build] Port 4edf35f11a9e
LLVM GN Syncbot [Tue, 26 Jan 2021 19:12:09 +0000 (19:12 +0000)]
[gn build] Port 4edf35f11a9e

3 years agoSupport for instrumenting only selected files or functions
Petr Hosek [Fri, 15 Jan 2021 09:14:37 +0000 (01:14 -0800)]
Support for instrumenting only selected files or functions

This change implements support for applying profile instrumentation
only to selected files or functions. The implementation uses the
sanitizer special case list format to select which files and functions
to instrument, and relies on the new noprofile IR attribute to exclude
functions from instrumentation.

Differential Revision: https://reviews.llvm.org/D94820

3 years ago[libomptarget][devicertl][amdgpu] Fix build, variable renaming error
Jon Chesterfield [Tue, 26 Jan 2021 19:05:21 +0000 (19:05 +0000)]
[libomptarget][devicertl][amdgpu] Fix build, variable renaming error

3 years ago[clangd] FindTarget resolves base specifier
Nathan James [Tue, 26 Jan 2021 18:59:29 +0000 (18:59 +0000)]
[clangd] FindTarget resolves base specifier

FindTarget on the virtual keyword or access specifier of a base specifier will now resolve to type of the base specifier.

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D95338

3 years ago[clangd] Selection handles CXXBaseSpecifier
Nathan James [Tue, 26 Jan 2021 18:58:53 +0000 (18:58 +0000)]
[clangd] Selection handles CXXBaseSpecifier

Selection now includes the virtual and access modifier as part of their range for cxx base specifiers.

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D95231

3 years ago[ARM] [ELF] Fix ARMMaterializeGV for Indirect calls
Adhemerval Zanella [Tue, 26 Jan 2021 18:20:26 +0000 (15:20 -0300)]
[ARM] [ELF] Fix ARMMaterializeGV for Indirect calls

Recent shouldAssumeDSOLocal changes (introduced by 961f31d8ad14c66)
do not take in consideration the relocation model anymore.  The ARM
fast-isel pass uses the function return to set whether a global symbol
is loaded indirectly or not, and without the expected information
llvm now generates an extra load for following code:

```
$ cat test.ll
@__asan_option_detect_stack_use_after_return = external global i32
define dso_local i32 @main(i32 %argc, i8** %argv) #0 {
entry:
  %0 = load i32, i32* @__asan_option_detect_stack_use_after_return,
align 4
  %1 = icmp ne i32 %0, 0
  br i1 %1, label %2, label %3

2:
  ret i32 0

3:
  ret i32 1
}

attributes #0 = { noinline optnone }

$ lcc test.ll -o -
[...]
main:
        .fnstart
[...]
        movw    r0, :lower16:__asan_option_detect_stack_use_after_return
        movt    r0, :upper16:__asan_option_detect_stack_use_after_return
        ldr     r0, [r0]
        ldr     r0, [r0]
        cmp     r0, #0
[...]
```

And without 'optnone' it produces:
```
[...]
main:
        .fnstart
[...]
        movw    r0, :lower16:__asan_option_detect_stack_use_after_return
        movt    r0, :upper16:__asan_option_detect_stack_use_after_return
        ldr     r0, [r0]
        clz     r0, r0
        lsr     r0, r0, #5
        bx      lr

[...]
```

This triggered a lot of invalid memory access in sanitizers for
arm-linux-gnueabihf.  I checked this patch both a stage1 built with
gcc and a stage2 bootstrap and it fixes all the Linux sanitizers
issues.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D95379

3 years ago[RISCV] Have customLegalizeToWOp truncate to the original type instead of i32 now...
Craig Topper [Tue, 26 Jan 2021 18:47:49 +0000 (10:47 -0800)]
[RISCV] Have customLegalizeToWOp truncate to the original type instead of i32 now that we use it for i8/i16 as well.

239cfbccb0509da1a08d9e746706013b732e646b add support for legalizing
i8/i16 UDIV/UREM/SDIV to use *W instructions. So we need to truncate
to i8/i16 if we're legalizing one of those.

3 years ago[mlir] sret and byval now require a type argument when constructed.
Eric Schweitz [Tue, 26 Jan 2021 18:18:44 +0000 (10:18 -0800)]
[mlir] sret and byval now require a type argument when constructed.

Fixes the LLVM code gen bugs and adds the missing tests.

Reviewed By: ftynse

Differential Revision: https://reviews.llvm.org/D95378

3 years agoReland "[lit] Use os.cpu_count() to cleanup TODO"
Julian Lettner [Fri, 15 Jan 2021 01:04:39 +0000 (17:04 -0800)]
Reland "[lit] Use os.cpu_count() to cleanup TODO"

The initial problem with the remaining bot config was resolved.

We can now use Python3.  Let's use `os.cpu_count()` to cleanup this
helper.

Differential Revision: https://reviews.llvm.org/D94734

3 years ago[lldb][NFC] Another attempt to fix GCC 5.x compilation
Raphael Isemann [Tue, 26 Jan 2021 17:13:45 +0000 (18:13 +0100)]
[lldb][NFC] Another attempt to fix GCC 5.x compilation

37510f69b4cb8d76064f108d57bebe95984a23ae tried to fix GCC 5.x compilation
by making the enum which is used as a unordered_map key unscoped. However it
seems that in GCC 5.x, enum keys are not supported *at all* in unordered_maps
(at least that's what some trial&error on godbolt tells me). This updates the
workaround to just use an int until GCC 5.x support is dropped.

3 years ago[mlir] Set CUDA/ROCm context before creating resources.
Christian Sigg [Tue, 26 Jan 2021 13:24:43 +0000 (14:24 +0100)]
[mlir] Set CUDA/ROCm context before creating resources.

The current context is thread-local state, and in preparation of GPU async execution (on multiple threads) we need to set the context before calling API that create resources.

Reviewed By: herhut

Differential Revision: https://reviews.llvm.org/D94495

3 years agoAMDGPU: Fix redundant FP spilling/assert in some functions
Matt Arsenault [Mon, 21 Sep 2020 12:43:06 +0000 (08:43 -0400)]
AMDGPU: Fix redundant FP spilling/assert in some functions

If a function has stack objects, and a call, we require an FP. If we
did not initially have any stack objects, and only introduced them
during PrologEpilogInserter for CSR VGPR spills, SILowerSGPRSpills
would end up spilling the FP register as if it were a normal
register. This would result in an assert in a debug build, or
redundant handling of the FP register in a release build.

Try to predict that we will have an FP later, although this is ugly.

3 years agoAMDGPU: Add assertion to determineCalleeSaves
Matt Arsenault [Thu, 21 Jan 2021 18:19:50 +0000 (13:19 -0500)]
AMDGPU: Add assertion to determineCalleeSaves

Make sure this isn't getting called multiple times. I was surprised we
were modifying the function here, which I think is a bit questionable.

3 years ago[OpenMP][deviceRTLs] Build the deviceRTLs with OpenMP instead of target dependent...
Shilei Tian [Tue, 26 Jan 2021 17:28:15 +0000 (12:28 -0500)]
[OpenMP][deviceRTLs] Build the deviceRTLs with OpenMP instead of target dependent language

From this patch (plus some landed patches), `deviceRTLs` is taken as a regular OpenMP program with just `declare target` regions. In this way, ideally, `deviceRTLs` can be written in OpenMP directly. No CUDA, no HIP anymore. (Well, AMD is still working on getting it work. For now AMDGCN still uses original way to compile) However, some target specific functions are still required, but they're no longer written in target specific language. For example, CUDA parts have all refined by replacing CUDA intrinsic and builtins with LLVM/Clang/NVVM intrinsics.
Here're a list of changes in this patch.
1. For NVPTX, `DEVICE` is defined empty in order to make the common parts still work with AMDGCN. Later once AMDGCN is also available, we will completely remove `DEVICE` or probably some other macros.
2. Shared variable is implemented with OpenMP allocator, which is defined in `allocator.h`. Again, this feature is not available on AMDGCN, so two macros are redefined properly.
3. CUDA header `cuda.h` is dropped in the source code. In order to deal with code difference in various CUDA versions, we build one bitcode library for each supported CUDA version. For each CUDA version, the highest PTX version it supports will be used, just as what we currently use for CUDA compilation.
4. Correspondingly, compiler driver is also updated to support CUDA version encoded in the name of bitcode library. Now the bitcode library for NVPTX is named as `libomptarget-nvptx-cuda_[cuda_version]-sm_[sm_number].bc`, such as `libomptarget-nvptx-cuda_80-sm_20.bc`.

With this change, there are also multiple features to be expected in the near future:
1. CUDA will be completely dropped when compiling OpenMP. By the time, we also build bitcode libraries for all supported SM, multiplied by all supported CUDA version.
2. Atomic operations used in `deviceRTLs` can be replaced by `omp atomic` if OpenMP 5.1 feature is fully supported. For now, the IR generated is totally wrong.
3. Target specific parts will be wrapped into `declare variant` with `isa` selector if it can work properly. No target specific macro is needed anymore.
4. (Maybe more...)

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D94745

3 years ago[lldb] Remove unused ThreadPlanStack::GetStackOfKind (NFC)
Dave Lee [Tue, 26 Jan 2021 00:52:32 +0000 (16:52 -0800)]
[lldb] Remove unused ThreadPlanStack::GetStackOfKind (NFC)

This function isn't used.

Differential Revision: https://reviews.llvm.org/D95411

3 years ago[clangd] Add std::size_t to StdSymbol mapping
Kadir Cetinkaya [Tue, 26 Jan 2021 06:39:43 +0000 (07:39 +0100)]
[clangd] Add std::size_t to StdSymbol mapping

This is a common symbol that's missing from our mapping because
cppreference yields multiple headers.

Add it manually by picking cstddef to prevent insertion of some stdlib-internal
headers instead.

Fixes https://github.com/clangd/clangd/issues/666.

Differential Revision: https://reviews.llvm.org/D95423

3 years ago[mlir] Add Python bindings for IntegerSet
Alex Zinenko [Mon, 25 Jan 2021 17:17:19 +0000 (18:17 +0100)]
[mlir] Add Python bindings for IntegerSet

This follows up on the introduction of C API for the same object and is similar
to AffineExpr and AffineMap.

Reviewed By: stellaraccident

Differential Revision: https://reviews.llvm.org/D95437

3 years ago[LoopVectorize] add test for fmin/fmax FMF propagation; NFC
Sanjay Patel [Tue, 26 Jan 2021 15:43:38 +0000 (10:43 -0500)]
[LoopVectorize] add test for fmin/fmax FMF propagation; NFC

The existing test has less FMF than we might expect if
our FMF was fixed (on all FP values), so this additional
test is intended to check propagation in a more "normal"
example.

3 years ago[LoopUtils] do not initialize Cmp predicate unnecessarily; NFC
Sanjay Patel [Tue, 26 Jan 2021 15:25:36 +0000 (10:25 -0500)]
[LoopUtils] do not initialize Cmp predicate unnecessarily; NFC

The switch must set the predicate correctly; anything else
should lead to unreachable/assert.

I'm trying to fix FMF propagation here and the callers,
so this is a preliminary cleanup.

3 years agoFix null dereference static analysis warning. NFCI.
Simon Pilgrim [Tue, 26 Jan 2021 16:19:18 +0000 (16:19 +0000)]
Fix null dereference static analysis warning. NFCI.

Replace cast_or_null<> with cast<> as we immediately dereference the pointer afterward so we're not expecting a null pointer.

3 years ago[AMDGPU] HSAMD::fromString - replace std::string arg with StringRef. NFCI.
Simon Pilgrim [Tue, 26 Jan 2021 16:09:27 +0000 (16:09 +0000)]
[AMDGPU] HSAMD::fromString - replace std::string arg with StringRef. NFCI.

Removes an unnecessary chain of StringRef -> std::string -> StringRef conversions

3 years ago[NFC] Disallow unused prefixes under clang/test/CodeGen
Mircea Trofin [Tue, 26 Jan 2021 04:44:41 +0000 (20:44 -0800)]
[NFC] Disallow unused prefixes under clang/test/CodeGen

Differential Revision: https://reviews.llvm.org/D95417

3 years ago[mlir][nfc] Move `getInnermostParallelLoops` to SCF/Transforms/Utils.h.
Alexander Belyaev [Tue, 26 Jan 2021 15:59:19 +0000 (16:59 +0100)]
[mlir][nfc] Move `getInnermostParallelLoops` to SCF/Transforms/Utils.h.

3 years ago[Sema] diagnoseEquivalentInternalLinkageDeclarations - assert for non-null NamedDecl...
Simon Pilgrim [Tue, 26 Jan 2021 15:51:49 +0000 (15:51 +0000)]
[Sema] diagnoseEquivalentInternalLinkageDeclarations - assert for non-null NamedDecl. NFCI.

Fixes clang static analysis warnings.

3 years ago[AMDGPU] Fix null-dereference static analysis warnings. NFCI.
Simon Pilgrim [Tue, 26 Jan 2021 15:33:31 +0000 (15:33 +0000)]
[AMDGPU] Fix null-dereference static analysis warnings. NFCI.

Avoid repeated calls to isZeroValue() and check for a null pointer before dereferencing a dyn_cast<>.

3 years ago[libomptarget][NFC] Fixed obsolete function names in comments
George Rokos [Tue, 26 Jan 2021 15:37:21 +0000 (07:37 -0800)]
[libomptarget][NFC] Fixed obsolete function names in comments

3 years agoAMDGPU: Clear IsSSA property in SIFormMemoryClauses
Matt Arsenault [Mon, 18 Jan 2021 15:49:48 +0000 (10:49 -0500)]
AMDGPU: Clear IsSSA property in SIFormMemoryClauses

Fixes verifier error when writing MIR testcases

3 years ago[LoopUnswitch] Avoid partially unswitching too aggressively.
Florian Hahn [Fri, 22 Jan 2021 22:49:45 +0000 (22:49 +0000)]
[LoopUnswitch] Avoid partially unswitching too aggressively.

This patch adds additional checks to avoid partial unswitching
in cases where it won't be profitable, e.g. because the path directly
exits the loop anyways.

3 years ago[LoopUnswitch] Add some additional tests.
Florian Hahn [Fri, 22 Jan 2021 22:47:43 +0000 (22:47 +0000)]
[LoopUnswitch] Add some additional tests.

Add a few additional tests where partial unswitching is not really
profitable and should be avoided.

3 years agoFix signed/unsigned comparison warning. NFCI.
Simon Pilgrim [Tue, 26 Jan 2021 14:57:37 +0000 (14:57 +0000)]
Fix signed/unsigned comparison warning. NFCI.

3 years ago[CostModel] Handle CTLZ and CCTZ in getTypeBasedIntrinsicInstrCost
Sander de Smalen [Tue, 26 Jan 2021 14:03:55 +0000 (14:03 +0000)]
[CostModel] Handle CTLZ and CCTZ in getTypeBasedIntrinsicInstrCost

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D95355

3 years ago[AMDGPU] Add IntrWillReturn to three intrinsics
Sebastian Neubauer [Tue, 26 Jan 2021 11:44:02 +0000 (12:44 +0100)]
[AMDGPU] Add IntrWillReturn to three intrinsics

None of these can terminate a wave or lane.
With these, all intrinsic are IntrWillReturn except those that change
exec or can terminate the wave.

Not marking intrinsics as WillReturn may prevent optimizations in the
future: https://lists.llvm.org/pipermail/llvm-dev/2021-January/148047.html

Differential Revision: https://reviews.llvm.org/D95436

3 years ago[AMDGPU] Fix use of HasModifiers in VopProfile
Mirko Brkusanin [Tue, 26 Jan 2021 14:21:11 +0000 (15:21 +0100)]
[AMDGPU] Fix use of HasModifiers in VopProfile

HasModifiers should be true if at least one modifier is used.
This should make the use of this field bit more consistent.

Differential Revision: https://reviews.llvm.org/D94795

3 years ago[Passes] Run peeling as part of simple/full loop unrolling.
Florian Hahn [Tue, 26 Jan 2021 13:43:39 +0000 (13:43 +0000)]
[Passes] Run peeling as part of simple/full loop unrolling.

Loop peeling removes conditions from loop bodies that become invariant
after a small number of iterations. When triggered, this leads to fewer
compares and possibly PHIs in loop bodies, enabling further
optimizations. The current cost-model of loop peeling should be quite
conservative/safe, i.e. only peel if a condition in the loop becomes
known after peeling.

For example, see PR47671, where loop peeling enables vectorization by
removing a PHI the vectorizer does not understand. Granted, the
loop-vectorizer could also be taught about constant PHIs, but loop
peeling is likely to enable other optimizations as well.

This has an impact on quite a few benchmarks from
MultiSource/SPEC2000/SPEC2006 on X86 with -O3 -flto, for example

    Same hash: 186 (filtered out)
    Remaining: 51
    Metric: loop-vectorize.LoopsVectorized

    Program                                        base   patch  diff
     test-suite...ve-susan/automotive-susan.test     8.00   9.00 12.5%
     test-suite...nal/skidmarks10/skidmarks.test    35.00  31.00 -11.4%
     test-suite...lications/sqlite3/sqlite3.test    41.00  43.00  4.9%
     test-suite...s/ASC_Sequoia/AMGmk/AMGmk.test    25.00  26.00  4.0%
     test-suite...006/450.soplex/450.soplex.test    88.00  89.00  1.1%
     test-suite...TimberWolfMC/timberwolfmc.test   120.00 119.00 -0.8%
     test-suite.../CINT2006/403.gcc/403.gcc.test   215.00 216.00  0.5%
     test-suite...006/447.dealII/447.dealII.test   957.00 958.00  0.1%
     test-suite...ternal/HMMER/hmmcalibrate.test    75.00  75.00  0.0%

    Same hash: 186 (filtered out)
    Remaining: 51
    Metric: loop-vectorize.LoopsAnalyzed

    Program                                        base    patch   diff
     test-suite...ks/Prolangs-C/agrep/agrep.test   440.00  434.00  -1.4%
     test-suite...nal/skidmarks10/skidmarks.test   312.00  308.00  -1.3%
     test-suite...marks/7zip/7zip-benchmark.test   6399.00 6323.00 -1.2%
     test-suite...lications/minisat/minisat.test   134.00  135.00   0.7%
     test-suite...rks/FreeBench/pifft/pifft.test   295.00  297.00   0.7%
     test-suite...TimberWolfMC/timberwolfmc.test   1879.00 1869.00 -0.5%
     test-suite...pplications/treecc/treecc.test   689.00  691.00   0.3%
     test-suite...T2000/300.twolf/300.twolf.test   1593.00 1597.00  0.3%
     test-suite.../Benchmarks/Bullet/bullet.test   1394.00 1392.00 -0.1%
     test-suite...ications/JM/ldecod/ldecod.test   1431.00 1429.00 -0.1%
     test-suite...6/464.h264ref/464.h264ref.test   2229.00 2230.00  0.0%
     test-suite...lications/sqlite3/sqlite3.test   2590.00 2589.00 -0.0%
     test-suite...ications/JM/lencod/lencod.test   2732.00 2733.00  0.0%
     test-suite...006/453.povray/453.povray.test   3395.00 3394.00 -0.0%

Note the -11% regression in number of loops vectorized for skidmarks. I
suspect this corresponds to the fact that those loops are gone now (see
the reduction in number of loops analyzed by LV).

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D88471

3 years ago[ORC] Attempt to auto-claim responsibility for weak defs in ObjectLinkingLayer.
Lang Hames [Tue, 26 Jan 2021 12:56:01 +0000 (23:56 +1100)]
[ORC] Attempt to auto-claim responsibility for weak defs in ObjectLinkingLayer.

Compilers may insert new definitions during compilation, E.g. EH personality
function pointers, or named constant pool entries. This commit causes
ObjectLinkingLayer to attempt to claim responsibility for all weak definitions
in objects as they're linked. This is always safe (first claimant for each
symbol is granted responsibility, subsequent claims are rejected without error)
and prevents compiler-injected symbols from being dead-stripped (which they
will be if they remain unclaimed by anyone).

This change was motivated by errors seen by an out-of-tree client while testing
eh-frame support in JITLink ELF/x86-64: IR containing exceptions didn't define
DW.ref.__gxx_personality_v0 (since it's added by CodeGen), and this caused
DW.ref.__gxx_personality_v0 to be dead-stripped leading to linker failures.

No test case yet: We won't have a way to test in-tree until we enable JITLink
for lli on Linux.

3 years agoRevert "[flang] Search for #include "file" in right directory"
Andrzej Warzynski [Tue, 26 Jan 2021 12:54:12 +0000 (12:54 +0000)]
Revert "[flang] Search for #include "file" in right directory"

This reverts commit d987b61b1dce9948801ac37704477e7c257100b1.

As pointed out in https://reviews.llvm.org/D95388, the reverted commit
causes build failures in the following Flang buildbots:
  * http://lab.llvm.org:8011/#/builders/32/builds/2642
  * http://lab.llvm.org:8011/#/builders/33/builds/2131
  * http://lab.llvm.org:8011/#/builders/135/builds/1473
  * http://lab.llvm.org:8011/#/builders/66/builds/1559
  * http://lab.llvm.org:8011/#/builders/134/builds/1409
  * http://lab.llvm.org:8011/#/builders/132/builds/1817
I'm guessing that the patch was only tested with
`FLANG_BUILD_NEW_DRIVER=Off` (i.e. the default). The builders listed
above set `FLANG_BUILD_NEW_DRIVER` to `On`.

Although fixing the build is relatively easy, the reverted patch
modifies the behaviour of the frontend, which breaks driver tests. In
particular, in https://reviews.llvm.org/D93453 support for `-I` was
added that depends on the current behaviour. The reverted patch
changes that behaviour. Either the tests have to be updated or the
change fine-tuned.