platform/upstream/llvm.git
2 years ago[AArch64][AsmParser] Arch directives should set implied features.
Sander de Smalen [Thu, 24 Feb 2022 08:46:15 +0000 (08:46 +0000)]
[AArch64][AsmParser] Arch directives should set implied features.

When assembling for example an SVE instruction with the `.arch +sve2` directive,
+sve should be implied by setting +sve2, similar to what would happen if
one would pass the mattr=+sve2 flag on the command-line.

The AsmParser doesn't set the implied features, meaning that the SVE
instruction does not assemble. This patch fixes that.

Note that the same does not hold when disabling a feature. For example,
+nosve2 does not imply +nosve.

Reviewed By: c-rhodes

Differential Revision: https://reviews.llvm.org/D120259

2 years ago[clang-format] Do not insert space after new/delete keywords in C function declarations
Luis Penagos [Thu, 24 Feb 2022 08:48:24 +0000 (09:48 +0100)]
[clang-format] Do not insert space after new/delete keywords in C function declarations

Fixes https://github.com/llvm/llvm-project/issues/46915.

Reviewed By: curdeius, HazardyKnusperkeks

Differential Revision: https://reviews.llvm.org/D120374

2 years ago[libc++] Remove <chrono> include from <future>
Nikolas Klauser [Wed, 23 Feb 2022 22:07:07 +0000 (23:07 +0100)]
[libc++] Remove <chrono> include from <future>

Reviewed By: Quuxplusone, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D120436

2 years ago[C++20][Modules][3/8] Initial handling for module partitions.
Iain Sandoe [Sun, 7 Feb 2021 01:00:33 +0000 (01:00 +0000)]
[C++20][Modules][3/8] Initial handling for module partitions.

This implements the parsing and recognition of module partition CMIs
and removes the FIXMEs in the parser.

Module partitions are recognised in the base computation of visibility,
however additional amendments to visibility follow in subsequent patches.

Differential Revision: https://reviews.llvm.org/D118586

2 years ago[RISCV] Add schedule class for Zbm and Zbe extension
lian wang [Tue, 15 Feb 2022 06:17:33 +0000 (06:17 +0000)]
[RISCV] Add schedule class for Zbm and Zbe extension

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D119805

2 years ago[RISCV][NFC] Remove useless intrinsic function declare in test of Zbp extension
lian wang [Thu, 24 Feb 2022 01:57:10 +0000 (01:57 +0000)]
[RISCV][NFC] Remove useless intrinsic function declare in test of Zbp extension

Reviewed By: benshi001

Differential Revision: https://reviews.llvm.org/D120393

2 years ago[libcxx] [test] Fix the get/put long_double_ru_RU tests on Apple platforms
Martin Storsjö [Tue, 22 Feb 2022 10:28:08 +0000 (12:28 +0200)]
[libcxx] [test] Fix the get/put long_double_ru_RU tests on Apple platforms

This fixes issue #45084 (https://llvm.org/PR45739).

Remove unnecessary trailing spaces after the "RUB" international
currency symbol (and after the plain number in some parts of the
put_long_double test).

Both of these test files are
`XFAIL: netbsd || linux || LIBCXX-WINDOWS-FIXME`, and then have some of
their test cases commented out when `__APPLE__`. This patch comments-in
those test cases and adjusts them all to work on Apple, while leaving the
test `XFAIL`ed on NetBSD, Linux, and Windows.

Differential Revision: https://reviews.llvm.org/D120316

2 years ago[analyzer] Fix -Wunused-lambda-capture in -DLLVM_ENABLE_ASSERTIONS=off builds
Fangrui Song [Thu, 24 Feb 2022 08:13:13 +0000 (00:13 -0800)]
[analyzer] Fix -Wunused-lambda-capture in -DLLVM_ENABLE_ASSERTIONS=off builds

2 years ago[InstCombine] Canonicalize SPF to min/max intrinsics
Nikita Popov [Mon, 14 Feb 2022 16:18:07 +0000 (17:18 +0100)]
[InstCombine] Canonicalize SPF to min/max intrinsics

Now that integer min/max intrinsics have good support in both
InstCombine and other passes, start canonicalizing SPF min/max
to intrinsic min/max.

Once this sticks, we can stop matching SPF min/max in various
places, and can remove hacks we have for preventing infinite loops
and breaking of SPF canonicalization.

Differential Revision: https://reviews.llvm.org/D98152

2 years agoRevert "[InstCombine] Remove one-use limitation from X-Y==0 fold"
Nikita Popov [Thu, 24 Feb 2022 07:50:40 +0000 (08:50 +0100)]
Revert "[InstCombine] Remove one-use limitation from X-Y==0 fold"

This reverts commit 65dc78d63ee2eb20fbed54401091f08a685ef8c1.

This caused a major code-size regression on tramp3d-v4, revert
until I can investigate.

2 years ago[mlir][SCF] Apply suggested ClangTidy fixes.
Adrian Kuegel [Thu, 24 Feb 2022 07:22:13 +0000 (08:22 +0100)]
[mlir][SCF] Apply suggested ClangTidy fixes.

2 years ago[BOLT] Support PC-relative relocations with addends
Maksim Panchenko [Thu, 24 Feb 2022 06:54:42 +0000 (22:54 -0800)]
[BOLT] Support PC-relative relocations with addends

PC-relative memory operand could reference a different object from
the one located at the target address, e.g. when a negative offset
is used. Check relocations for the real referenced object.

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D120379

2 years ago[Driver][RISCV] Add missing rv64 test case
eopXD [Tue, 22 Feb 2022 03:20:02 +0000 (19:20 -0800)]
[Driver][RISCV] Add missing rv64 test case

Add missing test case. Thanks @rogfer01 for spotting this out.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D120297

2 years ago[RISCV] Add more tests for vcpop and vfirst with VL=0
Chenbing.Zheng [Thu, 24 Feb 2022 05:23:25 +0000 (05:23 +0000)]
[RISCV] Add more tests for vcpop and vfirst with VL=0

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D120300

2 years ago[RISCV][NFC] Add helper function isVectorConfigInstr to reduce Repeated code.
Chenbing.Zheng [Thu, 24 Feb 2022 03:49:58 +0000 (03:49 +0000)]
[RISCV][NFC] Add helper function isVectorConfigInstr to reduce Repeated code.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D119924

2 years ago[ELF] Symbols.h: remove #include "InputFiles.h"
Fangrui Song [Thu, 24 Feb 2022 05:36:45 +0000 (21:36 -0800)]
[ELF] Symbols.h: remove #include "InputFiles.h"

2 years ago[ELF] Move isUsedInRegularObj assignment from ctor to call sites. NFC
Fangrui Song [Thu, 24 Feb 2022 05:32:50 +0000 (21:32 -0800)]
[ELF] Move isUsedInRegularObj assignment from ctor to call sites. NFC

This removes the tricky
`isUsedInRegularObj(!file || file->kind() == InputFile::ObjKind)`
and the copy from `Symbol::mergeProperties`.

2 years ago[RISCV] Use SelectionDAG::getFreeze to simplify some code. NFC
Craig Topper [Thu, 24 Feb 2022 05:12:54 +0000 (21:12 -0800)]
[RISCV] Use SelectionDAG::getFreeze to simplify some code. NFC

2 years ago[ELF][test] Avoid race on a.out
Fangrui Song [Thu, 24 Feb 2022 04:48:49 +0000 (20:48 -0800)]
[ELF][test] Avoid race on a.out

2 years ago[ELF] Don't rely on Symbols.h's transitive inclusion of InputFiles.h. NFC
Fangrui Song [Thu, 24 Feb 2022 04:44:34 +0000 (20:44 -0800)]
[ELF] Don't rely on Symbols.h's transitive inclusion of InputFiles.h. NFC

2 years ago[BOLT][NFC] Report errors from RewriteInstance `discoverStorage` and `run`
Amir Ayupov [Thu, 24 Feb 2022 03:30:30 +0000 (19:30 -0800)]
[BOLT][NFC] Report errors from RewriteInstance `discoverStorage` and `run`

Further improve error handling in BOLT by reporting `RewriteInstance` errors in
a library and fuzzer-friendly way instead of exiting.

Follow-up to D119658

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D120224

2 years ago[ELF] Move TLS mismatch error from Symbol::replace to postParse
Fangrui Song [Thu, 24 Feb 2022 04:34:48 +0000 (20:34 -0800)]
[ELF] Move TLS mismatch error from Symbol::replace to postParse

* detect `def_tls.o undef_nontls.o` violation
* place error checking code (checking duplicate symbol) together
* allow `--defsym tls1=tls2 def_tls.o`

As a degraded error checking, `--defsym tls1=42` violation will not be detected.

2 years ago[flang] Do not print format tabs
V Donaldson [Thu, 24 Feb 2022 00:31:07 +0000 (16:31 -0800)]
[flang] Do not print format tabs

As an extension, tabs are accepted in a format, but should be skipped,
not printed.

2 years ago[HIP] Support `-fgpu-default-stream`
Yaxun (Sam) Liu [Tue, 22 Feb 2022 01:16:04 +0000 (20:16 -0500)]
[HIP] Support `-fgpu-default-stream`

Introduce -fgpu-default-stream={legacy|per-thread} option to
support per-thread default stream for HIP runtime.

When -fgpu-default-stream=per-thread, HIP kernels are
launched through hipLaunchKernel_spt instead of
hipLaunchKernel. Also HIP_API_PER_THREAD_DEFAULT_STREAM=1
is defined by the preprocessor to enable other per-thread stream
API's.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D120298

2 years ago[ELF] Don't rely on Symbols.h's transitive inclusion of InputFiles.h. NFC
Fangrui Song [Thu, 24 Feb 2022 03:18:24 +0000 (19:18 -0800)]
[ELF] Don't rely on Symbols.h's transitive inclusion of InputFiles.h. NFC

2 years ago[mlir][sparse][taco] Add support for float32.
Bixia Zheng [Tue, 22 Feb 2022 22:04:30 +0000 (14:04 -0800)]
[mlir][sparse][taco] Add support for float32.

Previously, we only support float64. We now support float32 and float64. When
constructing a tensor without providing a data type, the default is float32.

Fix the tests to data type consistency. All PyTACO application tests now use
float32 to match the default data type of TACO. Other tests may use float32 or
float64.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D120356

2 years ago[ELF] Remove SharedSymbol::getFile. NFC
Fangrui Song [Thu, 24 Feb 2022 01:57:52 +0000 (17:57 -0800)]
[ELF] Remove SharedSymbol::getFile. NFC

Symbol.h depends on InputFiles.h. This change moves us toward dropping the
weird dependency.

The call sites will become slightly uglier (`cast<SharedFile>(s->file)`), but
the compromise is acceptable.

2 years ago[ScalarEvolution] Control flag for nonstrict inequalities in finite loops
Mircea Trofin [Thu, 24 Feb 2022 01:45:32 +0000 (17:45 -0800)]
[ScalarEvolution] Control flag for nonstrict inequalities in finite loops

D118090 causes a pretty significant (19%) regression in some Eigen
benchmarks. Investigating is a bit time consuming as the compilation
unit where this occurs is large. Rather than revert, this patch adds a
flag controlling that behavior (enabled by default).

2 years ago[mlir][sparse][linalg] add linalg rewriting specific to sparse tensors
Aart Bik [Thu, 24 Feb 2022 00:22:42 +0000 (16:22 -0800)]
[mlir][sparse][linalg] add linalg rewriting specific to sparse tensors

Now that sparse tensor types are first-class citizens and the sparse compiler
is taking shape, it is time to make sure other compiler optimizations compose
well with sparse tensors. Mostly, this should be completely transparent (i.e.,
dense and sparse take the same path). However, in some cases, optimizations
only make sense in the context of sparse tensors. This is a first example of
such an optimization, where fusing a sampled elt-wise multiplication only makes
sense when the resulting kernel has a potential lower asymptotic complexity due
to the sparsity.

As an extreme example, running SDDMM with 1024x1024 matrices and a sparse
sampling matrix with only two elements runs in 463.55ms in the unfused
case but just 0.032ms in the fused case, with a speedup of 14485x that
is only possible in the exciting world of sparse computations!

Reviewed By: mravishankar

Differential Revision: https://reviews.llvm.org/D120429

2 years ago[lldb/test] Fix TestProgressReporting.py race issue with the event listener
Med Ismail Bennani [Thu, 24 Feb 2022 00:42:53 +0000 (16:42 -0800)]
[lldb/test] Fix TestProgressReporting.py race issue with the event listener

This patch is a follow-up of D120100 to address some feedbacks from
@labath.

This should mainly fix the race issue with the even listener by moving
the listener setup to the main thread.

This also changes the SBDebugger::GetProgressFromEvent SWIG binding
arguments to be output only, so the user don't have to provide them.

Finally, this updates the test to check it the out arguments are returned
in a tuple and re-enables the test on all platforms.

Differential Revision: https://reviews.llvm.org/D120284

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years agoPGOInstrumentation, GCOVProfiling: Split indirectbr critical edges regardless of...
Matthias Braun [Wed, 16 Feb 2022 23:01:37 +0000 (15:01 -0800)]
PGOInstrumentation, GCOVProfiling: Split indirectbr critical edges regardless of PHIs

The `SplitIndirectBrCriticalEdges` function was originally designed for
`CodeGenPrepare` and skipped splitting of edges when the destination
block didn't contain any `PHI` instructions. This only makes sense when
reducing COPYs like `CodeGenPrepare`. In the case of
`PGOInstrumentation` or `GCOVProfiling` it would result in missed
counters and wrong result in functions with computed goto.

Differential Revision: https://reviews.llvm.org/D120096

2 years agoSimplify/cleanup BasicBlockUtilsTest
Matthias Braun [Thu, 17 Feb 2022 21:57:06 +0000 (13:57 -0800)]
Simplify/cleanup BasicBlockUtilsTest

Cleanup BasicBolckUtilsTest using C++ raw string literals, remove
duplicated block functions and smaller style changes.

Differential Revision: https://reviews.llvm.org/D120095

2 years ago[asan] Allow -fsanitize-address-globals-dead-stripping with -fno-data-sections for ELF
Fangrui Song [Thu, 24 Feb 2022 00:08:25 +0000 (16:08 -0800)]
[asan] Allow -fsanitize-address-globals-dead-stripping with -fno-data-sections for ELF

-fdata-sections decides whether global variables go into different sections.
This is orthogonal to whether we place their metadata (`.data` or `asan_globals`) into different sections.

With -fno-data-sections, `-fsanitize-address-globals-dead-stripping` can still:

* deduplicate COMDAT `asan.module_ctor` and `asan.module_dtor`
* (with ld --gc-sections): for a data section (e.g. `.data`), if all global variables defined relative to it are unreferenced, discard them and associated `asan_globals` sections (rare but no need to exclude this case)

Similar to c7b90947bd0179d914fea56b52be545c8f60f20a for PE/COFF.

Reviewed By: #sanitizers, kstoimenov, vitalybuka

Differential Revision: https://reviews.llvm.org/D120394

2 years ago[BOLT][NFC] Fix undefined behavior in encodeAnnotationImm
Amir Ayupov [Wed, 23 Feb 2022 06:54:15 +0000 (22:54 -0800)]
[BOLT][NFC] Fix undefined behavior in encodeAnnotationImm

Fix UBSan-reported issue in MCPlusBuilder::encodeAnnotationImm (left shift of a
negative value).

Test Plan:
```
ninja check-bolt
...
PASS: BOLT-Unit :: Core/./CoreTests/AArch64/MCPlusBuilderTester.Annotation/0 (1 of 140)
PASS: BOLT-Unit :: Core/./CoreTests/X86/MCPlusBuilderTester.Annotation/0 (131 of 134)
```

Reviewed By: maksfb, yota9

Differential Revision: https://reviews.llvm.org/D120260

2 years agoTeach the AArch64 backend to instruction select the BCAX instruction.
Owen Anderson [Fri, 18 Feb 2022 08:15:25 +0000 (00:15 -0800)]
Teach the AArch64 backend to instruction select the BCAX instruction.

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D120112

2 years agoReland "unbreak Modules/cxx20-export-import.cpp with LLVM_APPEND_VC_REV after 32b73bc...
Nico Weber [Wed, 23 Feb 2022 23:41:20 +0000 (18:41 -0500)]
Reland "unbreak Modules/cxx20-export-import.cpp with LLVM_APPEND_VC_REV after 32b73bc6ab82"

This reverts commit 5086cff04eec4327acc22a90466854ad4d89d570.
32b73bc6ab82 relanded in 1592d88aa7bc.

2 years ago[FormatVariadic] Mark index as required in docstring
Dave Lee [Wed, 2 Feb 2022 18:45:19 +0000 (10:45 -0800)]
[FormatVariadic] Mark index as required in docstring

After looking at the formatv docstring, I thought the index was optional (as it
is in other languages). This changes the header docs to show `index` instead of
`[index]`, to indicate that it is required.

Differential Revision: https://reviews.llvm.org/D118833

2 years ago[CMake] Use CMAKE_SYSROOT to build libs for Win to ARM cross tooolchain. NFC.
Vladimir Vereschaka [Thu, 17 Feb 2022 01:20:39 +0000 (17:20 -0800)]
[CMake] Use CMAKE_SYSROOT to build libs for Win to ARM cross tooolchain. NFC.

Provide CMAKE_SYSROOT for the libc++/libc++abi/libunwind libraries
instead of specific <foo>_SYSROOT for each of them.

Fixed passing some CMake arguments for the runtimes.

Referenced Differentials:
 * https://reviews.llvm.org/D119836
 * https://reviews.llvm.org/D112155
 * https://reviews.llvm.org/D111672

Differential Revision: https://reviews.llvm.org/D120383

2 years ago[libc++] Add empty line in ReleaseNotes.rst
Nikolas Klauser [Wed, 23 Feb 2022 23:09:18 +0000 (00:09 +0100)]
[libc++] Add empty line in ReleaseNotes.rst

2 years ago[OpenMP][NFC] Address warnings and lint messages in CGOpenMPRuntime
Joseph Huber [Wed, 23 Feb 2022 22:40:29 +0000 (17:40 -0500)]
[OpenMP][NFC] Address warnings and lint messages in CGOpenMPRuntime

Summary:
This patch addressed the warnings and linting messages for the
CGOpenMPRuntime.cpp file. This was causing some -Werror builds to fail.

2 years agoAdd support for floating-point option `ffp-eval-method` and for
Zahira Ammarguellat [Tue, 19 Oct 2021 16:12:57 +0000 (09:12 -0700)]
Add support for floating-point option `ffp-eval-method` and for
`pragma clang fp eval_method`.

Differential Revision: https://reviews.llvm.org/D109239

2 years ago[NFC][hwasan] Check _GLIBCXX_RELEASE in test
Vitaly Buka [Wed, 23 Feb 2022 22:29:09 +0000 (14:29 -0800)]
[NFC][hwasan] Check _GLIBCXX_RELEASE in test

Differential Revision: https://reviews.llvm.org/D119161

2 years ago[opt] Pin region viewer passes to legacy PM.
Michael Kruse [Tue, 22 Feb 2022 22:34:04 +0000 (16:34 -0600)]
[opt] Pin region viewer passes to legacy PM.

The RegionPrinter, RegionOnlyPrinter, RegionViewer and RegionOnlyViewer passes have not yet been ported to the new pass manager.

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D119897

2 years ago[HWASan] Use hwasan_memalign for aligned new.
Vitaly Buka [Wed, 23 Feb 2022 22:05:00 +0000 (14:05 -0800)]
[HWASan] Use hwasan_memalign for aligned new.

Aligned new does not require size to be a multiple of alignment, so
memalign is the correct choice instead of aligned_alloc.

Fixes false reports for unaligned sizes.

Reviewed By: eugenis

Differential Revision: https://reviews.llvm.org/D119161

2 years ago[SampleProf][Inliner] Add an option to turn off inliner in sample-profile pass.
minglotus-6 [Tue, 22 Feb 2022 20:07:43 +0000 (12:07 -0800)]
[SampleProf][Inliner] Add an option to turn off inliner in sample-profile pass.

Use case is offline evaluation (for inliner effectiveness) or debugging.

Differential Revision: https://reviews.llvm.org/D120344

2 years ago[NFC][hwasan] Clang-format the file
Vitaly Buka [Wed, 23 Feb 2022 22:04:51 +0000 (14:04 -0800)]
[NFC][hwasan] Clang-format the file

2 years agoUse function prototypes when appropriate; NFC
Aaron Ballman [Wed, 23 Feb 2022 22:11:34 +0000 (17:11 -0500)]
Use function prototypes when appropriate; NFC

2 years agoFix more unused lambda capture warnings, NFC
Reid Kleckner [Wed, 23 Feb 2022 22:05:26 +0000 (14:05 -0800)]
Fix more unused lambda capture warnings, NFC

2 years ago[libc++] Granularize chrono includes
Nikolas Klauser [Wed, 23 Feb 2022 22:05:22 +0000 (23:05 +0100)]
[libc++] Granularize chrono includes

Reviewed By: Quuxplusone, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D120141

2 years ago[HWASan] add test for debug info of allocas that don't need padding.
Florian Mayer [Tue, 15 Feb 2022 19:41:52 +0000 (11:41 -0800)]
[HWASan] add test for debug info of allocas that don't need padding.

Reviewed By: eugenis

Differential Revision: https://reviews.llvm.org/D119873

2 years agoFix unused lambda capture warning, NFC
Reid Kleckner [Wed, 23 Feb 2022 22:01:01 +0000 (14:01 -0800)]
Fix unused lambda capture warning, NFC

2 years ago[NVPTX] Add ex2.approx.f16/f16x2 support
Nicolas Miller [Tue, 22 Feb 2022 22:50:42 +0000 (14:50 -0800)]
[NVPTX] Add ex2.approx.f16/f16x2 support

his patch adds builtins and intrinsics for the f16 and f16x2 variants of the ex2
instruction.

These two variants were added in PTX7.0, and are supported by sm_75 and above.

Note that this isn't wired with the exp2 llvm intrinsic because the ex2
instruction is only available in its approx variant.

Running ptxas on the assembly generated by the test f16-ex2.ll works as
expected.

Differential Revision: https://reviews.llvm.org/D119157

2 years ago[NVPTX] Add more FMA intriniscs/builtins
Jakub Chlanda [Tue, 22 Feb 2022 22:45:19 +0000 (14:45 -0800)]
[NVPTX] Add more FMA intriniscs/builtins

This patch adds builtins/intrinsics for the following variants of FMA:

- f16, f16x2
  - rn
  - rn_ftz
  - rn_sat
  - rn_ftz_sat
  - rn_relu
  - rn_ftz_relu
- bf16, bf16x2
  - rn
  - rn_relu

ptxas (Cuda compilation tools, release 11.0, V11.0.194) is happy with the generated assembly.

Differential Revision: https://reviews.llvm.org/D118977

2 years ago[NVPTX] Expose float tys min, max, abs, neg as builtins
Jakub Chlanda [Tue, 22 Feb 2022 22:42:15 +0000 (14:42 -0800)]
[NVPTX] Expose float tys min, max, abs, neg as builtins

Adds support for the following builtins:

- abs, neg:
- .bf16,
- .bf16x2
- min, max
- {.ftz}{.NaN}{.xorsign.abs}.f16
- {.ftz}{.NaN}{.xorsign.abs}.f16x2
- {.NaN}{.xorsign.abs}.bf16
- {.NaN}{.xorsign.abs}.bf16x2
- {.ftz}{.NaN}{.xorsign.abs}.f32

Differential Revision: https://reviews.llvm.org/D117887

2 years ago[Clang][Docs] Add '-fopenmp-offload-mandatory' to command line reference
Joseph Huber [Wed, 23 Feb 2022 21:53:30 +0000 (16:53 -0500)]
[Clang][Docs] Add '-fopenmp-offload-mandatory' to command line reference

2 years ago[SLP] Fastpath instructions not in block being scheduled [nfc]
Philip Reames [Wed, 23 Feb 2022 21:48:03 +0000 (13:48 -0800)]
[SLP] Fastpath instructions not in block being scheduled [nfc]

2 years ago[OpenMP] Add option to make offloading mandatory
Joseph Huber [Tue, 22 Feb 2022 21:15:34 +0000 (16:15 -0500)]
[OpenMP] Add option to make offloading mandatory

Currently when we generate OpenMP offloading code we always make
fallback code for the CPU. This is necessary for implementing features
like conditional offloading and ensuring that unhandled pragmas don't
result in missing symbols. However, this is problematic for a few cases.
For offloading tests we can silently fail to the host without realizing
that offloading failed. Additionally, this makes it impossible to
provide interoperabiility to other offloading schemes like HIP or CUDA
because those methods do not provide any such host fallback guaruntee.
this patch adds the `-fopenmp-offload-mandatory` flag to prevent
generating the fallback symbol on the CPU and instead replaces the
function with a dummy global and the failed branch with 'unreachable'.

Reviewed By: ABataev

Differential Revision: https://reviews.llvm.org/D120353

2 years ago[SLP] Replace a impossible branch condition with an assert [NFC]
Philip Reames [Wed, 23 Feb 2022 21:42:49 +0000 (13:42 -0800)]
[SLP] Replace a impossible branch condition with an assert [NFC]

An entire bundle must be inside the scheduling window.  Assert that this property holds as opposed to checking it at runtime.

2 years ago[sanitizer][sancov] Use pc-1 for s390x
Fangrui Song [Wed, 23 Feb 2022 21:35:22 +0000 (13:35 -0800)]
[sanitizer][sancov] Use pc-1 for s390x

The stack trace addresses may be odd (normally addresses should be even), but
seems a good compromise when the instruction length (2,4,6) cannot be detected
easily.

Reviewed By: uweigand

Differential Revision: https://reviews.llvm.org/D120432

2 years ago[clang] Remove Address::deprecated() from CGClass.cpp
Arthur Eubanks [Wed, 23 Feb 2022 21:26:32 +0000 (13:26 -0800)]
[clang] Remove Address::deprecated() from CGClass.cpp

2 years ago{SLP] Make it clear ScheduleDataMap is keyed by instructions [NFC]
Philip Reames [Wed, 23 Feb 2022 21:28:29 +0000 (13:28 -0800)]
{SLP] Make it clear ScheduleDataMap is keyed by instructions [NFC]

2 years ago[ELF][test] Fix edata-etext.s
Fangrui Song [Wed, 23 Feb 2022 21:29:21 +0000 (13:29 -0800)]
[ELF][test] Fix edata-etext.s

2 years ago[instrprof] Rename the profile kind types to be more descriptive.
Snehasish Kumar [Thu, 17 Feb 2022 22:44:49 +0000 (14:44 -0800)]
[instrprof] Rename the profile kind types to be more descriptive.

Based on the discussion in D115393, I've updated the names to be more
descriptive.

Reviewed By: ellis, MaskRay

Differential Revision: https://reviews.llvm.org/D120092

2 years agoRevert "[SLP] Remove cap on schedule window size"
Philip Reames [Wed, 23 Feb 2022 21:08:18 +0000 (13:08 -0800)]
Revert "[SLP] Remove cap on schedule window size"

This reverts commit 6adf4b039e095224edbbecda5972e5e3353b53b6.  Reverting while investigating https://github.com/llvm/llvm-project/issues/54029

2 years agoRevert "[SLP] Simplify extendSchedulingRegion"
Philip Reames [Wed, 23 Feb 2022 21:08:10 +0000 (13:08 -0800)]
Revert "[SLP] Simplify extendSchedulingRegion"

This reverts commit 8c85f3a0523070ef656e30e368df0a679c1400cd.

2 years ago[OpenMP][Offloading] Change N back to 256 in bug49334.cpp
Shilei Tian [Wed, 23 Feb 2022 21:10:35 +0000 (16:10 -0500)]
[OpenMP][Offloading] Change N back to 256 in bug49334.cpp

2 years ago[libcxx] [test] Fix time.get.byname get_one for Glibc and Windows
Martin Storsjö [Tue, 25 Jan 2022 09:38:41 +0000 (09:38 +0000)]
[libcxx] [test] Fix time.get.byname get_one for Glibc and Windows

This matches the fixes for the wchar version in
f081cc50372f9415ef4fa2204a4b7f54153af455.

Differential Revision: https://reviews.llvm.org/D120283

2 years ago[DAGCombiner][TargetLowering] Pass SDValue by value to isMulAddWithConstProfitable.
Craig Topper [Wed, 23 Feb 2022 20:35:06 +0000 (12:35 -0800)]
[DAGCombiner][TargetLowering] Pass SDValue by value to isMulAddWithConstProfitable.

Internally to DAGCombiner the SDValues were passed by non-const
reference despite not being modified. They were then passed by
const reference to TLI.

This patch passes them by value which is consistent with the vast
majority of code.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D120420

2 years ago[DAGCombine] Extend combineCarryDiamond()
Pawe Bylica [Wed, 23 Feb 2022 18:26:48 +0000 (19:26 +0100)]
[DAGCombine] Extend combineCarryDiamond()

In combineCarryDiamond() use getAsCarry() to find more candidates for being a carry flag.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D118362

2 years ago[pseudo] fix an out-of-bound error in LRTable.
Haojian Wu [Wed, 23 Feb 2022 20:34:20 +0000 (21:34 +0100)]
[pseudo] fix an out-of-bound error in LRTable.

Fix window debug build.

2 years ago[lldb] Fix (unintentional) recursion in CommandObjectRegexCommand
Jonas Devlieghere [Wed, 23 Feb 2022 19:52:20 +0000 (11:52 -0800)]
[lldb] Fix (unintentional) recursion in CommandObjectRegexCommand

Jim noticed that the regex command is unintentionally recursive. Let's
use the following command regex as an example:

  (lldb) com regex humm 's/([^ ]+) ([^ ]+)/p %1 %2 %1 %2/'

If we call it with arguments foo bar, thing behave as expected:

  (lldb) humm foo bar
  (...)
  foo bar foo bar

However, if we include %2 in the arguments, things break down:

  (lldb) humm fo%2o bar
  (...)
  fobaro bar fobaro bar

The problem is that the implementation of the substitution is too naive.
It substitutes the %1 token into the target template in place, then does
the %2 substitution starting with the resultant string. So if the
previous substitution introduced a %2 token, it would get processed in
the second sweep, etc.

This patch addresses the issue by walking the command once and
substituting the % variables in place.

  (lldb) humm fo%2o bar
  (...)
  fo%2o bar fo%2o bar

Furthermore, this patch also reports an error if not enough variables
were provided and add support for substituting %0.

rdar://81236994

Differential revision: https://reviews.llvm.org/D120101

2 years ago[SLP] Rearrange fields in ScheduleData for density [NFC]
Philip Reames [Wed, 23 Feb 2022 19:57:56 +0000 (11:57 -0800)]
[SLP] Rearrange fields in ScheduleData for  density [NFC]

2 years ago[NFC][PowerPC] Fix the check-cpu.ll test case.
Stefan Pintilie [Wed, 23 Feb 2022 15:18:19 +0000 (09:18 -0600)]
[NFC][PowerPC] Fix the check-cpu.ll test case.

This test doesn't work because the CHECK-NOT line is actually checking
something that only exists on stderr and not stdout.
Changed the test so that we now check both stderr and stdout.
Changed the test so that we check pwr9, pwr10, and future. The cpu names of
power9 or power10 are not supported in the llc backend.

Reviewed By: nemanjai, #powerpc

Differential Revision: https://reviews.llvm.org/D120349

2 years ago[mlir] Add sectionMemoryMapper to ExecutionEngineOptions
Emilio Cota [Wed, 23 Feb 2022 15:51:36 +0000 (10:51 -0500)]
[mlir] Add sectionMemoryMapper to ExecutionEngineOptions

By specifying a sectionMemoryMapper, users can control how
memory for JIT code is allocated.

In particular, I need this in order to use a named memory
region so that profilers such as perf(1) can correctly label
execution cycles coming from JIT'ed code.

Reviewed-by: ezhulenev
Differential Revision: https://reviews.llvm.org/D120415

2 years ago[Driver] Add -fno-sanitize-address-globals-dead-stripping
Fangrui Song [Wed, 23 Feb 2022 19:51:30 +0000 (11:51 -0800)]
[Driver] Add -fno-sanitize-address-globals-dead-stripping

It's customary for these options to have the -fno- form which is sometimes
handy to work around issues. Using the supported driver option is preferred over
the internal cl::opt option `-mllvm -asan-globals-live-support=0`

Reviewed By: kstoimenov, vitalybuka

Differential Revision: https://reviews.llvm.org/D120391

2 years ago[SLP] Remove SchedulingPriority from ScheduleData [NFC]
Philip Reames [Wed, 23 Feb 2022 19:40:03 +0000 (11:40 -0800)]
[SLP] Remove SchedulingPriority from ScheduleData [NFC]

First step in trying to shrink the memory footprint of ScheduleData to improve cache locality.

2 years ago[PATCH] ASAN: Align declaration with definition of a fn
Martin Liska [Wed, 23 Feb 2022 19:25:17 +0000 (11:25 -0800)]
[PATCH] ASAN: Align declaration with definition of a fn

Fixes:
https://bugs.llvm.org/show_bug.cgi?id=51641

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D115447

2 years ago[TSan][NFC] fixup for comment of Shadow
Xu Mingjie [Wed, 23 Feb 2022 19:15:57 +0000 (11:15 -0800)]
[TSan][NFC] fixup for comment of Shadow

There should be 1-bit unused field between tid field and is_atomic field of Shadow.

Reviewed By: dvyukov, vitalybuka

Differential Revision: https://reviews.llvm.org/D119417

2 years agoRevert "[TSan][NFC] fixup for comment of Shadow"
Vitaly Buka [Wed, 23 Feb 2022 19:21:04 +0000 (11:21 -0800)]
Revert "[TSan][NFC] fixup for comment of Shadow"

Wrong author.

This reverts commit 6bff092e3ed4ae1f21b290f88cf7152cb331aa48.

2 years ago[SLP] Simplify extendSchedulingRegion
Philip Reames [Wed, 23 Feb 2022 16:51:33 +0000 (08:51 -0800)]
[SLP] Simplify extendSchedulingRegion

This change uses instruction's comesBefore method to simplify the code significantly. There's little compile time concern here because getSpillCost already calls comesBefore on every basic block which contains a vectorization candidate. The only additional times we'll build basic block ordering is when we can't schedule a vector candidate anywhere in the containing block.

Differential Revision: https://reviews.llvm.org/D120364

2 years ago[libc++][AIX] Fix trivial_abi return tests for unique_ptr/weak_ptr
Jinsong Ji [Wed, 23 Feb 2022 18:43:27 +0000 (13:43 -0500)]
[libc++][AIX] Fix trivial_abi return tests for unique_ptr/weak_ptr

The unique_ptr_ret and weak_ptr_ret tests are not expected to pass on
AIX. These tests check that unique_ptr and weak_ptr are returned by
value, but on AIX, all structs are always returned by reference.

```
3.9.6 Function Return Values
...

Note: Structures of any length and character strings longer than four
bytes are returned in a storage buffer allocated by the caller. The
address of this buffer is passed as a hidden first argument in GPR3,
which causes the first explicit argument word to be passed in GPR4. This
hidden argument is treated as a formal argument and corresponds to the
first word of the argument area.
```

Reviewed By: #powerpc, daltenty, #libc, Quuxplusone, philnik

Differential Revision: https://reviews.llvm.org/D119952

2 years agoAttributorAttributes: avoid a crashing on bad alignments
Augie Fackler [Fri, 11 Feb 2022 23:32:38 +0000 (18:32 -0500)]
AttributorAttributes: avoid a crashing on bad alignments

Prior to this change, LLVM would attempt to optimize an
aligned_alloc(33, ...) call to the stack. This flunked an assertion when
trying to emit the alloca, which crashed LLVM. Avoid that with extra
checks.

Differential Revision: https://reviews.llvm.org/D119604

2 years ago[MLIR][Presburger] PresburgerSet::subtract: automatically restore state on return
Arjun P [Tue, 22 Feb 2022 17:28:21 +0000 (17:28 +0000)]
[MLIR][Presburger] PresburgerSet::subtract: automatically restore state on return

Reviewed By: Groverkss

Differential Revision: https://reviews.llvm.org/D120339

2 years ago[TSan][NFC] fixup for comment of Shadow
Vitaly Buka [Wed, 23 Feb 2022 19:15:57 +0000 (11:15 -0800)]
[TSan][NFC] fixup for comment of Shadow

There should be 1-bit unused field between tid field and is_atomic field of Shadow.

Reviewed By: dvyukov, vitalybuka

Differential Revision: https://reviews.llvm.org/D119417

2 years ago[MLIR][Arith] Canonicalize cmpf(int to fp) to cmpi
William S. Moses [Wed, 16 Feb 2022 16:26:12 +0000 (11:26 -0500)]
[MLIR][Arith] Canonicalize cmpf(int to fp) to cmpi

Given a cmpf of either uitofp or sitofp and a constant, attempt to canonicalize it to a cmpi.

This PR rewrites equivalent code within LLVM to now apply to MLIR arith.

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D117257

2 years ago[flang] Lower function and subroutine calls
Valentin Clement [Wed, 23 Feb 2022 18:48:07 +0000 (19:48 +0100)]
[flang] Lower function and subroutine calls

This patch introduce basic function/subroutine calls.
Because of the state of lowering only simple scalar arguments
can be used in the calls. This will be enhanced in follow up
patches with arrays, allocatable, pointer ans so on.

```
subroutine sub1()
end

subroutine sub2()
  call sub1()
end
```

This patch is part of the upstreaming effort from fir-dev branch.

Reviewed By: schweitz

Differential Revision: https://reviews.llvm.org/D120419

Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: V Donaldson <vdonaldson@nvidia.com>
2 years ago[X86] combineX86ShufflesRecursively - pull out repeated getValueType/getSimpleValueTy...
Simon Pilgrim [Wed, 23 Feb 2022 18:45:17 +0000 (18:45 +0000)]
[X86] combineX86ShufflesRecursively - pull out repeated getValueType/getSimpleValueType calls.

2 years agoRevert "[MachineOutliner][AArch64] NFC: Split MBBs into "outlinable ranges""
Jessica Paquette [Wed, 23 Feb 2022 18:35:52 +0000 (10:35 -0800)]
Revert "[MachineOutliner][AArch64] NFC: Split MBBs into "outlinable ranges""

This reverts commit d97f997eb79d91b2872ac13619f49cb3a7120781.

This commit was not NFC.

(See: https://reviews.llvm.org/rGd97f997eb79d91b2872ac13619f49cb3a7120781)

2 years ago[mlir] Async: update condition for dispatching block-aligned compute function
Eugene Zhulenev [Thu, 17 Feb 2022 18:22:18 +0000 (10:22 -0800)]
[mlir] Async: update condition for dispatching block-aligned compute function

+ compare block size with the unrollable inner dimension
+ reduce nesting in the code and simplify a bit IR building

Reviewed By: cota

Differential Revision: https://reviews.llvm.org/D120075

2 years ago[ELF] Check COMMON symbols for PROVIDE and don't redefine COMMON symbols edata/end...
Fangrui Song [Wed, 23 Feb 2022 18:15:42 +0000 (10:15 -0800)]
[ELF] Check COMMON symbols for PROVIDE and don't redefine COMMON symbols edata/end/etext

In GNU ld, the definition precedence is: regular symbol assignment > relocatable object definition > `PROVIDE` symbol assignment.

GNU ld's internal linker scripts define the non-reserved (by C and C++)
edata/end/etext with `PROVIDE` so the relocatable object definition takes
precedence. This makes sense because `int end;` is valid.

We currently redefine such symbols if they are COMMON, but not if they are
regular definitions, so `int end;` with -fcommon is essentially a UB in ld.lld.
Fix this (also improve consistency and match GNU ld) by using the
`isDefined` code path for `isCommon`. In GNU ld, reserved identifiers like
`__ehdr_start` do not use `PROVIDE`, while we treat them all as `PROVIDE`, this
seems fine.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D120389

2 years ago[llvm-profgen] Support symbol loading for debug fission
wlei [Tue, 22 Feb 2022 20:09:01 +0000 (12:09 -0800)]
[llvm-profgen] Support symbol loading for debug fission

Support to load debug info from dwarf split file, like .dwo, .dwp files. Leverage the `getNonSkeletonUnitDIE(false)` API to achieve this.

Add test cause to make sure all the ranges is well retrieved by the loader.

Reviewed By: ayermolo, hoy, wenlei

Differential Revision: https://reviews.llvm.org/D115973

2 years ago[clang] Remove getPointerElementType() in EmitVTableTypeCheckedLoad()
Arthur Eubanks [Wed, 23 Feb 2022 17:37:53 +0000 (09:37 -0800)]
[clang] Remove getPointerElementType() in EmitVTableTypeCheckedLoad()

2 years ago[X86] combineX86ShufflesRecursively - don't both widening inputs before calling combi...
Simon Pilgrim [Wed, 23 Feb 2022 17:29:41 +0000 (17:29 +0000)]
[X86] combineX86ShufflesRecursively - don't both widening inputs before calling combineX86ShuffleChain

combineX86ShuffleChain no longer has to assume that the shuffle inputs are the right size, so don't create unnecessary nodes messing up oneuse limits as detailed on Issue #45319

2 years ago[gn build] Port 7c1ee5e95f31
LLVM GN Syncbot [Wed, 23 Feb 2022 17:12:13 +0000 (17:12 +0000)]
[gn build] Port 7c1ee5e95f31

2 years ago[DAG] try to convert multiply to shift via demanded bits
Sanjay Patel [Wed, 23 Feb 2022 16:25:01 +0000 (11:25 -0500)]
[DAG] try to convert multiply to shift via demanded bits

This is a fix for a regression discussed in:
https://github.com/llvm/llvm-project/issues/53829

We cleared more high multiplier bits with 995d400,
but that can lead to worse codegen because we would fail
to recognize the now disguised multiplication by neg-power-of-2
as a shift-left. The problem exists independently of the IR
change in the case that the multiply already had cleared high
bits. We also convert shl+sub into mul+add in instcombine's
negator.

This patch fills in the high-bits to see the shift transform
opportunity. Alive2 attempt to show correctness:
https://alive2.llvm.org/ce/z/GgSKVX

The AArch64, RISCV, and MIPS diffs look like clear wins. The
x86 code requires an extra move register in the minimal examples,
but it's still an improvement to get rid of the multiply on all
CPUs that I am aware of (because multiply is never as fast as a
shift).

There's a potential follow-up noted by the TODO comment. We
should already convert that pattern into shl+add in IR, so
it's probably not common:
https://alive2.llvm.org/ce/z/7QY_Ga

Fixes #53829

Differential Revision: https://reviews.llvm.org/D120216

2 years agoRevert "AttributorAttributes: avoid a crashing on bad alignments"
Arthur Eubanks [Wed, 23 Feb 2022 17:08:03 +0000 (09:08 -0800)]
Revert "AttributorAttributes: avoid a crashing on bad alignments"

This reverts commit 70ff6fbeb9b5acb4995dc42286954b762d0937fd.

Breaks bots, e.g. http://45.33.8.238/linux/69375/step_12.txt.

2 years ago[flang][NFC] Clean up ConvertType
Valentin Clement [Wed, 23 Feb 2022 17:04:15 +0000 (18:04 +0100)]
[flang][NFC] Clean up ConvertType

This patch removes unused or obsolete code in
the ConvertType.h and ConvertType.cpp files. These
files were landed together with the initial flang
upstreaming. This cleanup will help future upstreaming
effort from fir-dev and keep only used code.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D120405

2 years ago[MLIR][Pass] Have PassRegistryEntry own pass strings
Daniel Resnick [Tue, 22 Feb 2022 17:41:48 +0000 (10:41 -0700)]
[MLIR][Pass] Have PassRegistryEntry own pass strings

This eliminates the requirement that pass-related strings outlive pass
instances, which will facilitate future work enabling dynamic passes
written in other languages.

Differential Revision: https://reviews.llvm.org/D120341

2 years ago[flang] Lower complex constant
Valentin Clement [Wed, 23 Feb 2022 17:01:58 +0000 (18:01 +0100)]
[flang] Lower complex constant

Add ability to lower complex constant.

This patch is part of the upstreaming effort from fir-dev branch.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D120402

Co-authored-by: Kiran Chandramohan <kiran.chandramohan@arm.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
2 years ago[Pseudo] Token/TokenStream, PP directive parser.
Sam McCall [Mon, 7 Feb 2022 18:11:16 +0000 (19:11 +0100)]
[Pseudo] Token/TokenStream, PP directive parser.

The TokenStream class is the representation of the source code that will
be fed into the GLR parser.

This patch allows a "raw" TokenStream to be built by reading source code.
It also supports scanning a TokenStream to find the directive structure.

Next steps (with placeholders in the code): heuristically choosing a
path through #ifs, preprocessing the code by stripping directives and comments.
These will produce a suitable stream to feed into the parser proper.

Differential Revision: https://reviews.llvm.org/D119162