platform/upstream/llvm.git
2 years ago[lld-macho] Handle LOH_ARM64_ADRP_LDR linker optimization hints
Daniel Bertalan [Thu, 30 Jun 2022 09:01:18 +0000 (11:01 +0200)]
[lld-macho] Handle LOH_ARM64_ADRP_LDR linker optimization hints

This linker optimization hint transforms a pair of adrp+ldr (immediate)
instructions into an ldr (literal) load from a PC-relative address if
it is 4-byte aligned and within +/- 1 MiB, as ldr can encode a signed
19-bit offset that gets multiplied by 4.

In the wild, only a small number of these hints are applicable because
not many loads end up close enough to the data segment. However, the
added helper functions will be useful in implementing the rest of the
LOH types.

Differential Revision: https://reviews.llvm.org/D128942

2 years ago[FPEnv] Allow CompoundStmt to keep FP options
Serge Pavlov [Mon, 28 Sep 2020 07:32:06 +0000 (14:32 +0700)]
[FPEnv] Allow CompoundStmt to keep FP options

AST does not have special nodes for pragmas. Instead a pragma modifies
some state variables of Sema, which in turn results in modified
attributes of AST nodes. This technique applies to floating point
operations as well. Every AST node that can depend on FP options keeps
current set of them.

This technique works well for options like exception behavior or fast
math options. They represent instructions to the compiler how to modify
code generation for the affected nodes. However treatment of FP control
modes has problems with this technique. Modifying FP control mode
(like rounding direction) usually requires operations on hardware, like
writing to control registers. It must be done prior to the first
operation that depends on the control mode. In particular, such
operations are required for implementation of `pragma STDC FENV_ROUND`,
compiler should set up necessary rounding direction at the beginning of
compound statement where the pragma occurs. As there is no representation
for pragmas in AST, the code generation becomes a complicated task in
this case.

To solve this issue FP options are kept inside CompoundStmt. Unlike to FP
options in expressions, these does not affect any operation on FP values,
but only inform the codegen about the FP options that act in the body of
the statement. As all pragmas that modify FP environment may occurs only
at the start of compound statement or at global level, such solution
works for all relevant pragmas. The options are kept as a difference
from the options in the enclosing compound statement or default options,
it helps codegen to set only changed control modes.

Differential Revision: https://reviews.llvm.org/D123952

2 years ago[ConstExpr] Don't create insertvalue expressions
Nikita Popov [Wed, 29 Jun 2022 08:48:40 +0000 (10:48 +0200)]
[ConstExpr] Don't create insertvalue expressions

In preparation for the removal in D128719, this stops creating
insertvalue constant expressions (well, unless they are directly
used in LLVM IR).

Differential Revision: https://reviews.llvm.org/D128792

2 years ago[mlir][SCF] Add a ParallelCombiningOpInterface to decouple scf::PerformConcurrently...
Nicolas Vasilache [Thu, 30 Jun 2022 10:37:21 +0000 (03:37 -0700)]
[mlir][SCF] Add a ParallelCombiningOpInterface to decouple scf::PerformConcurrently from its contained operations

This allows purging references of scf.ForeachThreadOp and scf.PerformConcurrentlyOp from
ParallelInsertSliceOp.
This will allowmoving the op closer to tensor::InsertSliceOp with which it should share much more
code.

In the future, the decoupling will also allow extending the type of ops that can be used in the
parallel combinator as well as semantics related to multiple concurrent inserts to the same
result.

Differential Revision: https://reviews.llvm.org/D128857

2 years ago[mlir][vector] Untangle TransferWriteDistribution and avoid crashing in the 0-D case.
Nicolas Vasilache [Wed, 29 Jun 2022 08:59:33 +0000 (01:59 -0700)]
[mlir][vector] Untangle TransferWriteDistribution and avoid crashing in the 0-D case.

This revision avoids a crash in the 0-D case of distributing vector.transfer ops out of
vector.warp_execute_on_lane_0.
Due to the code complexity and lack of documentation, it took untangling the implementation
before realizing that the simple fix was to fail in the 0-D case.
The rewrite is still very useful to understand this code better.

Differential Revision: https://reviews.llvm.org/D128793

2 years ago[SCCP] Only handle unknown lattice values in resolvedUndefsIn()
Nikita Popov [Tue, 21 Jun 2022 08:34:41 +0000 (10:34 +0200)]
[SCCP] Only handle unknown lattice values in resolvedUndefsIn()

This is a minor refinement of resolvedUndefsIn(), mostly for clarity.
If the value of an instruction is undef, then that's already a legal
final result -- we can safely rauw such an instruction with undef.
We only need to mark unknown values as overdefined, as that's the
result we get for an instruction that has not been processed because
it has an undef operand.

Differential Revision: https://reviews.llvm.org/D128251

2 years ago[AMDGPU] Add WMMA clang builtins
Piotr Sobczak [Fri, 1 Jul 2022 06:18:09 +0000 (08:18 +0200)]
[AMDGPU] Add WMMA clang builtins

Add WMMA clang builtins and tests. Extra changes in code
are needed to handle function overloads.

WavefrontSize 32:
__builtin_amdgcn_wmma_f32_16x16x16_f16_w32
__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32
__builtin_amdgcn_wmma_f16_16x16x16_f16_w32
__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32
__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32
__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32

WavefrontSize 64:
__builtin_amdgcn_wmma_f32_16x16x16_f16_w64
__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64
__builtin_amdgcn_wmma_f16_16x16x16_f16_w64
__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64
__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64
__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D128952

2 years ago[AMDGPU] Update WMMA intrinsics with explicit f16 types
Piotr Sobczak [Fri, 1 Jul 2022 06:18:03 +0000 (08:18 +0200)]
[AMDGPU] Update WMMA intrinsics with explicit f16 types

Update intrinsics to use n x f16 and n x i16 instead
of 32-bit types. This may avoid the need for a bitcast
and is probably less confusing.

Depends on making v16f16 and v16i16 types legal.

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D128951

2 years ago[NFC] add --match-full-lines to the RUN line
Chen Zheng [Fri, 1 Jul 2022 06:36:23 +0000 (02:36 -0400)]
[NFC] add --match-full-lines to the RUN line

2 years ago[pseudo] Define a clangPseudoCLI library.
Haojian Wu [Tue, 28 Jun 2022 20:37:03 +0000 (22:37 +0200)]
[pseudo] Define a clangPseudoCLI library.

- define a common data structure Language which is a compiled result of the
  bnf grammar. It is defined in Language.h;
- creates a clangPseudoCLI lib which defines a grammar commandline flag and
  expose a function to get the Language. It supports --grammar=cxx,
  --grammmar=/path/to/file.bnf;
- use the clangPseudoCLI in clang-pseudo, fuzzer, and benchmark tools (
  simplify the code and use the prebuilt cxx grammar);

Split out from https://reviews.llvm.org/D127448.

Differential Revision: https://reviews.llvm.org/D128679

2 years ago[flang] Fix APFloat conversion cases
Valentin Clement [Fri, 1 Jul 2022 06:29:19 +0000 (08:29 +0200)]
[flang] Fix APFloat conversion cases

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

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D128935

Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Peter Steinfeld <psteinfeld@nvidia.com>
2 years ago[Inline] don't add noalias metadata for unknown objects.
Chen Zheng [Fri, 24 Jun 2022 12:04:06 +0000 (08:04 -0400)]
[Inline] don't add noalias metadata for unknown objects.

The unidentified objects recognized in `getUnderlyingObjects` may
still alias to the noalias parameter because `getUnderlyingObjects`
may not check deep enough to get the underlying object because of
`MaxLookup`. The real underlying object for the unidentified object
 may still be the noalias parameter.

Originally Patched By: tingwang

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D127202

2 years ago[lldb] Add tests which simulate the various std::string layouts
Pavel Labath [Thu, 21 Apr 2022 08:57:18 +0000 (10:57 +0200)]
[lldb] Add tests which simulate the various std::string layouts

Checking whether a formatter change does not break some of the supported
string layouts is difficult because it requires tracking down and/or
building different versions and build configurations of the library.

The purpose of this patch is to avoid that by providing an in-tree
simulation of the string class. It is a reduced version of the real
string class, obtained by elimitating all non-trivial code, leaving
just the internal data structures used by the data formatter. Different
versions of the class can be simulated through preprocessor defines.

The test (ab)uses the fact that our formatters kick in for any
double-underscore sub-namespace of `std`, so it avoids colliding with
the real string class by declaring the test class in the std::__lldb
namespace.

I do not consider this to be a replacement for the existing data
formatter tests, as producing this kind of a test is not trivial, and it
is easy to make a mistake in the process. However, it's also not
realistic to expect that every person changing the data formatter will
test it against all versions of the real class, so I think it can be
useful as a first line of defence.

Adding support for new layouts can become particularly unwieldy, but
this complexity will also be reflected in the actual code, so if we find
ourselves needing to support too many variants, we may need to start
dropping support for old ones, or come up with a completely different
strategy.

Differential Revision: https://reviews.llvm.org/D124155

2 years ago[lldb/dyld-posix] Avoid reading the module list in inconsistent states
Pavel Labath [Mon, 20 Jun 2022 13:47:27 +0000 (15:47 +0200)]
[lldb/dyld-posix] Avoid reading the module list in inconsistent states

New glibc versions (since 2.34 or including this
<https://github.com/bminor/glibc/commit/ed3ce71f5c64c5f07cbde0ef03554ea8950d8f2c>
patch) trigger the rendezvous breakpoint after they have already added
some modules to the list. This did not play well with our dynamic
loader plugin which was doing a diff of the the reported modules in the
before (RT_ADD) and after (RT_CONSISTENT) states. Specifically, it
caused us to miss some of the modules.

While I think the old behavior makes more sense, I don't think that lldb
is doing the right thing either, as the documentation states that we
should not be expecting a consistent view in the RT_ADD (and RT_DELETE)
states.

Therefore, this patch changes the lldb algorithm to compare the module
list against the previous consistent snapshot. This fixes the previous
issue, and I believe it is more correct in general. It also reduces the
number of times we are fetching the module info, which should speed up
the debugging of processes with many shared libraries.

The change in RefreshModules ensures we don't broadcast the loaded
notification for the dynamic loader (ld.so) module more than once.

Differential Revision: https://reviews.llvm.org/D128264

2 years ago[PATCH] [lldb-server] Skip shared regions for memory allocation
Emre Kultursay [Fri, 1 Jul 2022 05:44:01 +0000 (13:44 +0800)]
[PATCH] [lldb-server] Skip shared regions for memory allocation

Differential Revision: https://reviews.llvm.org/D128832

2 years ago[clang][NFC][tests] dr208.c optional signext handling
Hubert Tong [Fri, 1 Jul 2022 04:03:58 +0000 (00:03 -0400)]
[clang][NFC][tests] dr208.c optional signext handling

Fixes llvm/llvm-project#56325.

2 years ago[mlir][Vector] Fold InsertStridedSliceOp of ExtractStridedSliceOp.
jacquesguan [Thu, 30 Jun 2022 11:24:31 +0000 (19:24 +0800)]
[mlir][Vector] Fold InsertStridedSliceOp of ExtractStridedSliceOp.

This patch supports to fold InsertStridedSliceOp(ExtractStridedSliceOp(dst), dst) to dst.

Differential Revision: https://reviews.llvm.org/D128903

2 years ago[mlir][Vector] Fold InsertStridedSliceOp of two splat with the same input to splat.
jacquesguan [Thu, 30 Jun 2022 08:30:59 +0000 (16:30 +0800)]
[mlir][Vector] Fold InsertStridedSliceOp of two splat with the same input to splat.

This patch folds InsertStridedSliceOp(SplatOp(X):src_type, SplatOp(X):dst_type) to SplatOp(X):dst_type.

Reviewed By: Mogball

Differential Revision: https://reviews.llvm.org/D128891

2 years ago[ODRHash diagnostics] Move common code for calculating diag locations in `DiagnoseODR...
Volodymyr Sapsai [Fri, 24 Jun 2022 02:05:50 +0000 (19:05 -0700)]
[ODRHash diagnostics] Move common code for calculating diag locations in `DiagnoseODRMismatch` into a lambda. NFC.

Differential Revision: https://reviews.llvm.org/D128489

2 years agoRemove unneeded cl::ZeroOrMore. NFC
Fangrui Song [Fri, 1 Jul 2022 02:11:27 +0000 (19:11 -0700)]
Remove unneeded cl::ZeroOrMore. NFC

2 years ago[mlir] Remove unneeded cl::ZeroOrMore for ListOption variables. NFC
Fangrui Song [Fri, 1 Jul 2022 02:04:44 +0000 (19:04 -0700)]
[mlir] Remove unneeded cl::ZeroOrMore for ListOption variables. NFC

2 years ago[ODRHash diagnostics] Split `err_module_odr_violation_mismatch_decl_diff` into per...
Volodymyr Sapsai [Thu, 23 Jun 2022 03:18:42 +0000 (20:18 -0700)]
[ODRHash diagnostics] Split `err_module_odr_violation_mismatch_decl_diff` into per-entity diagnostics. NFC.

We'll need to add more cases for Objective-C entities and adding
everything to `err_module_odr_violation_mismatch_decl_diff` makes it
harder to work with over time.

Differential Revision: https://reviews.llvm.org/D128488

2 years ago[mlir][tblgen] Improving error messages
wren romano [Mon, 27 Jun 2022 23:28:45 +0000 (16:28 -0700)]
[mlir][tblgen] Improving error messages

This differential improves two error conditions, by detecting them earlier and by providing better messages to help users understand what went wrong.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D128555

2 years ago[ISel] Match all bits when merge undefs for DAG combine
Xiang1 Zhang [Fri, 1 Jul 2022 01:05:36 +0000 (09:05 +0800)]
[ISel] Match all bits when merge undefs for DAG combine

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D128570

2 years agoRevert "[ISel] Match all bits when merge undef(s) for DAG combine"
Xiang1 Zhang [Fri, 1 Jul 2022 00:59:04 +0000 (08:59 +0800)]
Revert "[ISel] Match all bits when merge undef(s) for DAG combine"

This reverts commit 5fe5aa284efed1ee1492e1f266351b35f0a8bb69.

2 years ago[ISel] Match all bits when merge undef(s) for DAG combine
Xiang1 Zhang [Thu, 30 Jun 2022 11:07:25 +0000 (19:07 +0800)]
[ISel] Match all bits when merge undef(s) for DAG combine

2 years ago[SLP][NFC]Cleanup up operands of the removed insertelements, NFC.
Alexey Bataev [Fri, 1 Jul 2022 00:10:04 +0000 (17:10 -0700)]
[SLP][NFC]Cleanup up operands of the removed insertelements, NFC.

Replace all operands of the insertelement instruction, replaced by
shuffles, by poisons to avoid false-positive reports about incorrect function.

2 years ago[X86] Pre-commit tests for D128769. NFC
Craig Topper [Tue, 28 Jun 2022 23:10:22 +0000 (16:10 -0700)]
[X86] Pre-commit tests for D128769. NFC

2 years ago[RISCV] Avoid repeated code in SelectAddrRegImm. NFC
Craig Topper [Fri, 1 Jul 2022 00:15:55 +0000 (17:15 -0700)]
[RISCV] Avoid repeated code in SelectAddrRegImm. NFC

2 years ago[SVE] Use CPY to zero active lanes of a floating point vector.
Paul Walker [Fri, 24 Jun 2022 08:21:28 +0000 (09:21 +0100)]
[SVE] Use CPY to zero active lanes of a floating point vector.

Patterns exist for the integer case that are trivially expandable
to cover 0.0f.

Differential Revision: https://reviews.llvm.org/D128669

2 years ago[SVE] Extend "and(ipg,cmp(x,y))" patterns to cover the case when y is an immediate.
Paul Walker [Wed, 22 Jun 2022 15:05:17 +0000 (16:05 +0100)]
[SVE] Extend "and(ipg,cmp(x,y))" patterns to cover the case when y is an immediate.

Differential Revision: https://reviews.llvm.org/D128479

2 years ago[BOLT][DWARF] Support mix mode DWARF
Alexander Yermolovich [Thu, 30 Jun 2022 23:50:49 +0000 (16:50 -0700)]
[BOLT][DWARF] Support mix mode DWARF

Added support for mixing monolithic DWARF5 with legacy DWARF, and monolithic legacy and DWARF5 split dwarf.

Reviewed By: maksfb

Differential Revision: https://reviews.llvm.org/D128232

2 years ago[runtimes] adds llvm-libgcc to the list of runtimes to be sorted
Christopher Di Bella [Sat, 25 Jun 2022 00:17:30 +0000 (00:17 +0000)]
[runtimes] adds llvm-libgcc to the list of runtimes to be sorted

llvm-libgcc is not a part of `LLVM_ALL_RUNTIMES` because llvm-libgcc is
incompatible with an explicit libunwind and compiler-rt. This meant that
it was being filtered out and not built.

Differential Revision: https://reviews.llvm.org/D128568

2 years ago[MC][Mips] Support .reloc *, BFD_RELOC_{NONE,16,32,64}, *
Fangrui Song [Thu, 30 Jun 2022 23:39:23 +0000 (16:39 -0700)]
[MC][Mips] Support .reloc *, BFD_RELOC_{NONE,16,32,64}, *

... to match most other common architectures which already support BFD_RELOC_*.
BFD_RELOC_NONE provides a generic way indicating a dependency between two
sections and is useful for some instrumentations which encode symbol index
information (e.g. `.cg_profile`).

2 years ago[VE] Support load/store vm regsiters
Kazushi (Jam) Marukawa [Sat, 25 Jun 2022 02:34:08 +0000 (11:34 +0900)]
[VE] Support load/store vm regsiters

Support load/store vm registers to memory location as a first step.
As a next step, support load/store vm registers to stack location.
This patch also adds several regression tests for not only load/store
vm registers but also missing load/store for vr registers.

Reviewed By: efocht

Differential Revision: https://reviews.llvm.org/D128610

2 years ago[Sanitizer][Darwin] Lookup dyld image header via shared cache
Julian Lettner [Thu, 30 Jun 2022 19:13:15 +0000 (12:13 -0700)]
[Sanitizer][Darwin] Lookup dyld image header via shared cache

On macOS 13+, dyld itself has moved into the shared cache.  Looking it
up via vm_region_recurse_64() now causes spins/hangs/crashes.  We use a
different set of dyld APIs to find the image header in the shared cache.

rdar://92131949

Differential Revision: https://reviews.llvm.org/D128936

2 years agoAdds AST matcher for ObjCStringLiteral
Rashmi Mudduluru [Wed, 29 Jun 2022 21:21:42 +0000 (14:21 -0700)]
Adds AST matcher for ObjCStringLiteral

Differential Revision: https://reviews.llvm.org/D128103

2 years ago[RISCV] Remove an unnecessary copy of X0 in selectShiftMask.
Craig Topper [Thu, 30 Jun 2022 22:10:31 +0000 (15:10 -0700)]
[RISCV] Remove an unnecessary copy of X0 in selectShiftMask.

We know which instruction we're emitting so its ok to directly
encode X0 into the instruction. We only need to create a copy when
a constant 0 is selected without context of what instructions uses it.

2 years ago[NFC] Switch a few uses of undef to poison as placeholders for unreachble code
Nuno Lopes [Thu, 30 Jun 2022 22:01:27 +0000 (23:01 +0100)]
[NFC] Switch a few uses of undef to poison as placeholders for unreachble code

2 years agoImprove the formatting of static_assert messages
Corentin Jabot [Wed, 29 Jun 2022 17:13:19 +0000 (19:13 +0200)]
Improve the formatting of static_assert messages

Display 'static_assert failed: message' instead of
'static_assert failed "message"' to be consistent
with other implementations and be slightly more
readable.

Reviewed By: #libc, aaron.ballman, philnik, Mordante

Differential Revision: https://reviews.llvm.org/D128844

2 years ago[fix/build] Fix bazel build rule.
rdzhabarov [Thu, 30 Jun 2022 21:43:09 +0000 (21:43 +0000)]
[fix/build] Fix bazel build rule.

2 years ago[LLDB][NativePDB] Return LLDB_INVALID_ADDRESS in PdbIndex::MakeVirtualAddress when...
Zequan Wu [Thu, 30 Jun 2022 21:32:31 +0000 (14:32 -0700)]
[LLDB][NativePDB] Return LLDB_INVALID_ADDRESS in PdbIndex::MakeVirtualAddress when input is invalid due to missing address info in symbol/public records.

2 years ago[flang] Expand semantics test coverage of collective subroutines
Naje George [Thu, 30 Jun 2022 21:26:54 +0000 (14:26 -0700)]
[flang] Expand semantics test coverage of collective subroutines

Add non-standard conforming calls that violate the intent(inout)
of errmsg argument for co_sum, co_max, co_min, and co_broadcast.
Add non-standard conforming calls that violate the argument
typing of errmsg argument for co_max, co_min, and co_broadcast.
Add standard conforming calls that reorder keyword arguments
for co_sum and co_reduce.

Reviewed By: ktras

Differential Revision: https://reviews.llvm.org/D128468

2 years ago[InstCombine] Changing constant-indexed GEP of GEP to i8* for merging
William Huang [Thu, 30 Jun 2022 18:12:20 +0000 (18:12 +0000)]
[InstCombine] Changing constant-indexed GEP of GEP to i8* for merging

When merging GEP of GEP with constant indices, if the second GEP's offset is not divisible by the first GEP's element size, convert both type to i8* and merge.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D125934

2 years ago[RISCV] Make custom isel for (add X, imm) used by load/stores more selective.
Craig Topper [Thu, 30 Jun 2022 21:11:09 +0000 (14:11 -0700)]
[RISCV] Make custom isel for (add X, imm) used by load/stores more selective.

Only handle immediates that would produce an ADDI or ADDIW of Lo12
as the final instruction in their materialization.

As the test change show this removes immediates that materialize
with lui+addiw that is not the same as lui+addi.

2 years ago[gn build] Port 0f94d2b385e7
LLVM GN Syncbot [Thu, 30 Jun 2022 21:13:26 +0000 (21:13 +0000)]
[gn build] Port 0f94d2b385e7

2 years ago[AMDGPU] Make v16i16/v16f16 legal
Piotr Sobczak [Thu, 30 Jun 2022 19:58:57 +0000 (21:58 +0200)]
[AMDGPU] Make v16i16/v16f16 legal

There are upcoming intrinsics to use the new types.

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D128865

2 years agoFix test expectation positioning relative to FIXME comment
Reid Kleckner [Thu, 30 Jun 2022 21:07:46 +0000 (14:07 -0700)]
Fix test expectation positioning relative to FIXME comment

This new test was failing because the line number delta wasn't right.

2 years ago[flang] Add new semantics test for team_number function
Naje George [Thu, 30 Jun 2022 20:59:14 +0000 (13:59 -0700)]
[flang] Add new semantics test for team_number function

Add new semantics test for team_number and rename existing
team_number semantics test.

Reviewed By: ktras

Differential Revision: https://reviews.llvm.org/D128309

2 years ago[mlir] Add Dead Code Analysis
Mogball [Thu, 23 Jun 2022 19:02:45 +0000 (19:02 +0000)]
[mlir] Add Dead Code Analysis

This patch implements the analysis state classes needed for sparse data-flow analysis and implements a dead-code analysis using those states to determine liveness of blocks, control-flow edges, region predecessors, and function callsites.

Depends on D126751

Reviewed By: rriddle, phisiart

Differential Revision: https://reviews.llvm.org/D127064

2 years ago[NFC] Switch a few uses of undef to poison as placeholders for unreachble code
Nuno Lopes [Thu, 30 Jun 2022 20:46:19 +0000 (21:46 +0100)]
[NFC] Switch a few uses of undef to poison as placeholders for unreachble code

2 years ago[Kaleidoscope] Remove unused function argument
Marc Auberer [Thu, 30 Jun 2022 20:46:25 +0000 (20:46 +0000)]
[Kaleidoscope] Remove unused function argument

Removes an unused function argument from a code listing in the Kaleidoscope turorial in step 9.

Reviewed By: dblaikie, MaskRay

Differential Revision: https://reviews.llvm.org/D128628

2 years ago[libFuzzer] Extend the fuzz target intarface to allow -1 return value.
Kostya Serebryany [Tue, 28 Jun 2022 18:36:30 +0000 (11:36 -0700)]
[libFuzzer] Extend the fuzz target intarface to allow -1 return value.

With this change, fuzz targets may choose to return -1
to indicate that the input should not be added to the corpus
regardless of the coverage it generated.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D128749

2 years ago[InstCombine] Fix a Wparentheses warning in an assert. NFC
Craig Topper [Thu, 30 Jun 2022 20:03:32 +0000 (13:03 -0700)]
[InstCombine] Fix a Wparentheses warning in an assert. NFC

2 years ago[RISCV] Replace some uses of XLenVT in RISCVDAGToDAGISel::Select with the original...
Craig Topper [Thu, 30 Jun 2022 19:27:46 +0000 (12:27 -0700)]
[RISCV] Replace some uses of XLenVT in RISCVDAGToDAGISel::Select with the original Node VT. NFCI

These should contain the same thing, but we aren't consistent about
which we use.

Since we call ReplaceNode, it seems more correct to use the initial VT.

2 years ago[AMDGPU] GFX11: automatically release VGPRs at the end of the shader
Jay Foad [Thu, 30 Jun 2022 15:31:49 +0000 (16:31 +0100)]
[AMDGPU] GFX11: automatically release VGPRs at the end of the shader

GFX11 has a new message type MSG_DEALLOC_VGPRS which can be used to
release a shader's VGPRs. Sending this at the end of a shader (just
before the s_endpgm) can help overall system performance in cases where
the s_endpgm would have to wait for outstanding VMEM stores to complete
before releasing the VGPRs.

Differential Revision: https://reviews.llvm.org/D128442

2 years ago[InstCombine] canonicalize 'icmp (trunc X), C' to 'icmp (X & Mask), C'
Sanjay Patel [Thu, 30 Jun 2022 19:00:12 +0000 (15:00 -0400)]
[InstCombine] canonicalize 'icmp (trunc X), C' to 'icmp (X & Mask), C'

I looked at canonicalizing in the other direction, but that causes
many potential regressions and infinite loops because we already
(possibly wrongly) canonicalize "trunc X to i1" into an and+icmp.

This has a data layout restriction to avoid creating illegal
mask instructions, but we could remove that if we can show
that the backend can undo this when needed.

The motivating example from issue #56119 is modeled by the
PhaseOrdering test.

2 years ago[InstCombine] add tests for icmp (trunc X), C; NFC
Sanjay Patel [Thu, 30 Jun 2022 18:06:19 +0000 (14:06 -0400)]
[InstCombine] add tests for icmp (trunc X), C; NFC

2 years ago[PhaseOrdering] add test to show missing folds from PR56119; NFC
Sanjay Patel [Thu, 30 Jun 2022 17:58:23 +0000 (13:58 -0400)]
[PhaseOrdering] add test to show missing folds from PR56119; NFC

issue #56119

2 years ago[AMDGPU] add alternate tests for max-offset codegen; NFC
Sanjay Patel [Thu, 30 Jun 2022 12:10:51 +0000 (08:10 -0400)]
[AMDGPU] add alternate tests for max-offset codegen; NFC

As discussed in D128123, the existing test shows a possible
regression when converting sub to xor. This adds tests that
avoid that pattern but still has a offset near 65535. Also,
add a test with the canonical IR for the existing test to show
if the transform is happening with the expected pattern in IR.

2 years agoTest and document some C99 DRs
Aaron Ballman [Thu, 30 Jun 2022 19:46:13 +0000 (15:46 -0400)]
Test and document some C99 DRs

This captures the first 15 or so DRs in C99

2 years ago[SLP][X86] Add 32-bit vector stores to help vectorization opportunities
Simon Pilgrim [Thu, 30 Jun 2022 19:25:41 +0000 (20:25 +0100)]
[SLP][X86] Add 32-bit vector stores to help vectorization opportunities

Building on the work on D124284, this patch tags v4i8 and v2i16 vector loads as custom, enabling SLP to try to vectorize these types ending in a partial store (using the SSE MOVD instruction) - we already do something similar for 64-bit vector types.

Differential Revision: https://reviews.llvm.org/D127604

2 years ago[Interpreter] Pass target features to JIT
Jonas Hahnfeld [Wed, 29 Jun 2022 20:30:56 +0000 (22:30 +0200)]
[Interpreter] Pass target features to JIT

This is required to support RISC-V where the '+d' target feature
indicates the presence of the D instruction set extension, which
changes to the Hard-float 'd' ABI.

Differential Revision: https://reviews.llvm.org/D128853

2 years ago[pseudo] Add ForestNode descendants iterator, print ambiguous/opaque node stats.
Sam McCall [Thu, 30 Jun 2022 17:50:40 +0000 (19:50 +0200)]
[pseudo] Add ForestNode descendants iterator, print ambiguous/opaque node stats.

Differential Revision: https://reviews.llvm.org/D128930

2 years agoApparently you need a special makefile flag to use threads on Linux.
Jim Ingham [Thu, 30 Jun 2022 19:09:31 +0000 (12:09 -0700)]
Apparently you need a special makefile flag to use threads on Linux.

This is a follow-up to https://reviews.llvm.org/D128776.

2 years ago[OpenMP] Initial parsing and sema support for 'parallel masked taskloop' construct
Fazlay Rabbi [Thu, 30 Jun 2022 17:59:33 +0000 (10:59 -0700)]
[OpenMP] Initial parsing and sema support for 'parallel masked taskloop' construct

This patch gives basic parsing and semantic support for
"parallel masked taskloop" construct introduced in
OpenMP 5.1 (section 2.16.9)

Differential Revision: https://reviews.llvm.org/D128834

2 years agoThreads which hit a breakpoint but fail the condition are considered
Jim Ingham [Wed, 29 Jun 2022 01:06:30 +0000 (18:06 -0700)]
Threads which hit a breakpoint but fail the condition are considered
not to be hit.  But another thread might be hit at the same time and
actually stop.  So we have to be sure to switch the first thread's
stop info to eStopReasonNone or we'll report a hit when the condition
failed, which is confusing.

Differential Revision: https://reviews.llvm.org/D128776

2 years agoRevert "Deferred Concept Instantiation Implementation"
Jonas Devlieghere [Thu, 30 Jun 2022 18:35:12 +0000 (11:35 -0700)]
Revert "Deferred Concept Instantiation Implementation"

This reverts commit 2f207439521d62d9551b2884158368e8b34084e5 because it
triggers an assertion when building an LLDB test program:

  Assertion failed: (InstantiatingSpecializations.empty() && "failed to
  clean up an InstantiatingTemplate?"), function ~Sema, file
  /Users/buildslave/jenkins/workspace/lldb-cmake/llvm-project/clang/lib/Sema/Sema.cpp,
  line 458.

More details in https://reviews.llvm.org/D126907.

2 years ago[RISCV] Use getVTList to simplify creation of vleff MachineSDNode. NFC
Craig Topper [Thu, 30 Jun 2022 18:28:02 +0000 (11:28 -0700)]
[RISCV] Use getVTList to simplify creation of vleff MachineSDNode. NFC

We don't need to pass the 3 VTs separately, we already have a list
available to us.

2 years ago[llvm-objdump] Default to --mcpu=future for PPC64
Fangrui Song [Thu, 30 Jun 2022 18:30:34 +0000 (11:30 -0700)]
[llvm-objdump] Default to --mcpu=future for PPC64

GNU objdump disassembles all unknown instructions by default. Match this user
friendly behavior with the cpu value `future`.

Differential Revision: https://reviews.llvm.org/D127824

2 years ago[llvm-objdump] Default to --mattr=+all for AArch64
Fangrui Song [Thu, 30 Jun 2022 18:17:55 +0000 (11:17 -0700)]
[llvm-objdump] Default to --mattr=+all for AArch64

GNU objdump disassembles all unknown instructions by default. Match this user
friendly behavior with the target feature "all" (D128029) designed for disassemblers.

Reviewed By: jhenderson

Differential Revision: https://reviews.llvm.org/D128030

2 years ago[RISCV] Use the VT passed into selectImm instead of XLenVT. NFCI
Craig Topper [Thu, 30 Jun 2022 18:05:18 +0000 (11:05 -0700)]
[RISCV] Use the VT passed into selectImm instead of XLenVT. NFCI

I think the VT pased in will always be XLen.

2 years ago[flang] Add semantics test for `get_team`
Katherine Rasmussen [Wed, 22 Jun 2022 19:01:04 +0000 (12:01 -0700)]
[flang] Add semantics test for `get_team`

Add a semantics test for the intrinsic function `get_team`.

Differential Revision: https://reviews.llvm.org/D128370

2 years ago[mlir][Inliner] Support recursion in Inliner
Javed Absar [Sun, 5 Jun 2022 12:42:13 +0000 (13:42 +0100)]
[mlir][Inliner] Support recursion in Inliner

This fixes  Bug https://github.com/llvm/llvm-project/issues/53492
 and uses InlineHistory to track recursive inlining.

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D127072

2 years agoUseful error when input dim is unused by LHS/RHS.
Benoit Jacob [Thu, 30 Jun 2022 15:56:15 +0000 (15:56 +0000)]
Useful error when input dim is unused by LHS/RHS.

Differential Revision: https://reviews.llvm.org/D128925

2 years agoRevert "[clang] Emit SARIF Diagnostics: Create `clang::SarifDocumentWriter` interface"
Aaron Ballman [Thu, 30 Jun 2022 17:38:42 +0000 (13:38 -0400)]
Revert "[clang] Emit SARIF Diagnostics: Create `clang::SarifDocumentWriter` interface"

This reverts commit 329fae7103d355e728cc326a0a9abef889ccc577.

This should fix failing test bots like:
https://lab.llvm.org/buildbot/#/builders/91/builds/11328

2 years ago[AArch64] Add target feature "all"
Fangrui Song [Thu, 30 Jun 2022 17:37:58 +0000 (10:37 -0700)]
[AArch64] Add target feature "all"

This is used by disassemblers: `llvm-mc -disassemble -mattr=` and `llvm-objdump --mattr=`.
The main use case is for llvm-objdump to disassemble all known instructions
(D128030).

In user-facing tools, "all" is intentionally not supported in producers:
integrated assembler (`.arch_extension all`), clang -march (`-march=armv9.3a+all`).
Due to the code structure, `llvm-mc -mattr=+all` `llc -mattr=+all` are not
rejected (they are internal tool). Add `llvm/test/CodeGen/AArch64/mattr-all.ll`
to catch behavior changes.

AArch64SysReg::SysReg::haveFeatures: check `FeatureAll` to print
`AArch64SysReg::SysReg::AltName` for some system registers (e.g. `ERRIDR_EL1, RNDR`).

AArch64.td: add `AssemblerPredicateWithAll` to additionally test `FeatureAll`.
Change all `AssemblerPredicate` (except `UseNegativeImmediates`) to `AssemblerPredicateWithAll`.

utils/TableGen/{DecoderEmitter,SubtargetFeatureInfo}.cpp: support arbitrarily
nested all_of, any_of, and not.

Note: A predicate supports all_of, any_of, and not. For a target (though
currently not for AArch64) an encoding may be disassembled differently with
different target features.
Note: AArch64MCCodeEmitter::computeAvailableFeatures is not available to
the disassembler.

Reviewed By: peter.smith, lenary

Differential Revision: https://reviews.llvm.org/D128029

2 years ago[InstCombine] Fix memrchr logic error that prevents folding
Martin Sebor [Thu, 30 Jun 2022 17:22:28 +0000 (11:22 -0600)]
[InstCombine] Fix memrchr logic error that prevents folding

Correct a logic bug in the memrchr enhancement added in D123629 that
makes it ineffective in a subset of cases.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D128856

2 years ago[AArch64] Make FeatureFuseAdrpAdd a tune feature
Fangrui Song [Thu, 30 Jun 2022 17:32:38 +0000 (10:32 -0700)]
[AArch64] Make FeatureFuseAdrpAdd a tune feature

Update D120104 to add FeatureFuseAdrpAdd to Processor#TuneFeatures
instead of Processor#Features, similar to FeatureFuseAES, and matching
Tune*.

This enables FeatureFuseAdrpAdd for `clang -mcpu=xxx -mtune=generic` even
if xxx does not set FeatureFuseAdrpAdd.

Reviewed By: alexander-shaposhnikov, peter.smith

Differential Revision: https://reviews.llvm.org/D128787

2 years ago[clang] Emit SARIF Diagnostics: Create `clang::SarifDocumentWriter` interface
Vaibhav Yenamandra [Thu, 30 Jun 2022 17:23:15 +0000 (13:23 -0400)]
[clang] Emit SARIF Diagnostics: Create `clang::SarifDocumentWriter` interface

Create an interface for writing SARIF documents from within clang:

The primary intent of this change is to introduce the interface
clang::SarifDocumentWriter, which allows incrementally adding
diagnostic data to a JSON backed document. The proposed interface is
not yet connected to the compiler internals, which will be covered in
future work. As such this change will not change the input/output
interface of clang.

This change also introduces the clang::FullSourceRange type that is
modeled after clang::SourceRange + clang::FullSourceLoc, this is useful
for packaging a pair of clang::SourceLocation objects with their
corresponding SourceManagers.

Previous discussions:

RFC for this change: https://lists.llvm.org/pipermail/cfe-dev/2021-March/067907.html
https://lists.llvm.org/pipermail/cfe-dev/2021-July/068480.html
SARIF Standard (2.1.0):

https://docs.oasis-open.org/sarif/sarif/v2.1.0/os/sarif-v2.1.0-os.html

Differential Revision: https://reviews.llvm.org/D109701

2 years ago[llvm-reduce] Change initialization order to fix bots. NFC
Matthew Voss [Thu, 30 Jun 2022 16:44:23 +0000 (09:44 -0700)]
[llvm-reduce] Change initialization order to fix bots. NFC

Fixes this error:
/var/lib/buildbot/sanitizer-buildbot6/sanitizer-x86_64-linux-android/build/
llvm-project/llvm/tools/llvm-reduce/TestRunner.cpp:20:7:
error: field 'TM' will be initialized after field 'ToolName'
[-Werror,-Wreorder-ctor]
      TM(std::move(TM)), ToolName(ToolName) {
      ^~~~~~~~~~~~~~~~~  ~~~~~~~~~~~~~~~~~~
      Program(std::move(Program)) TM(std::move(TM))
1 error generated.

https://lab.llvm.org/buildbot/\#/builders/77/builds/19154

2 years ago[llvm-dwarfdump] --show-sources option to show all sources
Daniel Thornburgh [Wed, 16 Feb 2022 00:46:54 +0000 (00:46 +0000)]
[llvm-dwarfdump] --show-sources option to show all sources

This option allows printing all sources used by an object file.

Reviewed By: dblaikie, jhenderson

Differential Revision: https://reviews.llvm.org/D87656

2 years ago[SystemZ] Add support for tune-cpu attribute
Kai Nacke [Thu, 30 Jun 2022 15:29:44 +0000 (11:29 -0400)]
[SystemZ] Add support for tune-cpu attribute

clang (like gcc) has the `-mtune=` command line option. This option
adds the `"tune-cpu"` attribute to a function. The intended functionality
is that the scheduling model of that cpu is used. E.g. `-mtune=z15 -march=z14`
generates only instructions supported on z14 but uses the scheduling model
of z15 for it.
This PR adds the infrastructure to support this.

Reviewed By: uweigand

Differential Revision: https://reviews.llvm.org/D128910

2 years ago[mlir][VectorToGPU] Fix support for i4, col-major operand support
Christopher Bate [Fri, 17 Jun 2022 17:40:28 +0000 (11:40 -0600)]
[mlir][VectorToGPU] Fix support for i4, col-major operand support

For the conversion to nvgpu `mma.sync` and `ldmatrix` pathways, the code
was missing support for the `i4` data type. While fixing this, another
bug was discoverd that caused the number of ldmatrix tiles calculated for
certain operand types and configurations to be incorrect. This change
fixes both issues and adds additional tests.

Differential Revision: https://reviews.llvm.org/D128074

2 years ago[lldb] XFAIL TestObjCXXBridgedPO on macOS Ventura
Jonas Devlieghere [Thu, 30 Jun 2022 16:18:31 +0000 (09:18 -0700)]
[lldb] XFAIL TestObjCXXBridgedPO on macOS Ventura

TestObjCXXBridgedPO is broken on macOS Ventura (but not on macOS
Monterey). I took a look but it doesn't seem trivial. I'm XFAILing the
test until Adrian, who wrote the test, can take a look.

rdar://96224141

2 years ago[AMDGPU] Check for CopyToReg PhysReg clobbers in pre-RA-sched
jeff [Mon, 27 Jun 2022 20:17:00 +0000 (20:17 +0000)]
[AMDGPU] Check for CopyToReg PhysReg clobbers in pre-RA-sched

Differential Revision: https://reviews.llvm.org/D128681

2 years ago[flang] Fix "not yet implemented" message for CHARACTER MIN/MAX
Valentin Clement [Thu, 30 Jun 2022 16:10:53 +0000 (18:10 +0200)]
[flang] Fix "not yet implemented" message for CHARACTER MIN/MAX

The check to see if the arguments for the MIN/MAX intrinsics were of CHARACTER
type was not handling assumed length characters.  In this case, the FIR type is
"!fir.ref<!fir.char<1,?>>".

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

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D128922

Co-authored-by: Peter Steinfeld <psteinfeld@nvidia.com>
2 years ago[clang][dataflow] Replace TEST_F with TEST where possible
Sam Estep [Thu, 30 Jun 2022 16:03:23 +0000 (16:03 +0000)]
[clang][dataflow] Replace TEST_F with TEST where possible

Many of our tests are currently written using `TEST_F` where the test fixture class doesn't have any `SetUp` or `TearDown` methods, and just one helper method. In those cases, this patch deletes the class and pulls its method out into a standalone function, using `TEST` instead of `TEST_F`.

There are still a few test files leftover in `clang/unittests/Analysis/FlowSensitive/` that use `TEST_F`:

- `DataflowAnalysisContextTest.cpp` because the class contains a `Context` field which is used
- `DataflowEnvironmentTest.cpp` because the class contains an `Environment` field which is used
- `SolverTest.cpp` because the class contains a `Vals` field which is used
- `TypeErasedDataflowAnalysisTest.cpp` because there are several different classes which all share the same method name

Reviewed By: ymandel, sgatev

Differential Revision: https://reviews.llvm.org/D128924

2 years ago[RISCV] Fold (sra (add (shl X, 32), C1), 32 - C) -> (shl (sext_inreg (add X, C1), C)
Craig Topper [Thu, 30 Jun 2022 15:52:57 +0000 (08:52 -0700)]
[RISCV] Fold (sra (add (shl X, 32), C1), 32 - C) -> (shl (sext_inreg (add X, C1), C)

Similar for a subtract with a constant left hand side.

(sra (add (shl X, 32), C1<<32), 32) is the canonical IR from InstCombine
for (sext (add (trunc X to i32), 32) to i32).

For RISCV, we should lower this as addiw which means turning it into
(sext_inreg (add X, C1)).

There is an existing DAG combine to convert back to (sext (add (trunc X
to i32), 32) to i32), but it requires isTruncateFree to return true
and for i32 to be a legal type as it used sign_extend and truncate
nodes. So that doesn't work for RISCV.

If the outer sra happens be used by a shl by constant, it will be
folded and the shift amount of the sra will be changed before we
can do our own DAG combine. This requires us to match the more
general pattern and restore the shl.

I had wanted to do this as a separate (add (shl X, 32), C1<<32) ->
(shl (add X, C1), 32) combine, but that hit an infinite loop for some
values of C1.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D128869

2 years ago[RISCV] DAG combine (sra (shl X, 32), 32 - C) -> (shl (sext_inreg X, i32), C).
Craig Topper [Thu, 30 Jun 2022 15:52:43 +0000 (08:52 -0700)]
[RISCV] DAG combine (sra (shl X, 32), 32 - C) -> (shl (sext_inreg X, i32), C).

The sext_inreg can often be folded into an earlier instruction by
using a W instruction. The sext_inreg also works better with our ABI.

This is one of the steps to improving the generated code for this https://godbolt.org/z/hssn6sPco

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D128843

2 years ago[RISCV] Pre-commit tests for D128869. NFC
Craig Topper [Wed, 29 Jun 2022 23:23:41 +0000 (16:23 -0700)]
[RISCV] Pre-commit tests for D128869. NFC

2 years ago[llvm] Fix the modules build
Jonas Devlieghere [Thu, 30 Jun 2022 15:55:51 +0000 (08:55 -0700)]
[llvm] Fix the modules build

Fixes error: missing '#include "llvm/IR/FMF.h"'; 'FastMathFlags' must be
defined before it is used in llvm/include/llvm/IR/NoFolder.h.

2 years ago[llvm-reduce] Add support for LTO bitcode files
Matthew Voss [Thu, 30 Jun 2022 15:53:00 +0000 (08:53 -0700)]
[llvm-reduce] Add support for LTO bitcode files

Adds support for reading and writing LTO bitcode files.

  - Emit a summary if the original bitcode file had a summary
  - Use split LTO units if the original bitcode file used them.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D127168

2 years ago[flang] Fix one corner case in reshape intrinsic
Peixin Qiao [Thu, 30 Jun 2022 15:52:44 +0000 (23:52 +0800)]
[flang] Fix one corner case in reshape intrinsic

As Fortran 2018 16.9.163, the reshape is the only intrinsic which
requires the shape argument to be rank-one integer array and the SIZE
of it to be one constant expression. The current expression lowering
converts the shape expression with slice in intrinsic into one box value
with the box element type of unknown extent. However, the genReshape
requires the box element type to be constant size. So, convert the box
value into one with box element type of sequence of 1 x constant. This
corner case is found in cam4 in SPEC 2017
https://github.com/llvm/llvm-project/issues/56140.

Reviewed By: Jean Perier

Differential Revision: https://reviews.llvm.org/D128597

2 years ago[ARM] Add Thumb-1 CTTZ codegen tests. NFC
David Green [Thu, 30 Jun 2022 15:45:00 +0000 (16:45 +0100)]
[ARM] Add Thumb-1 CTTZ codegen tests. NFC

2 years ago[AMDGPU] gfx11 WMMA instruction support
Piotr Sobczak [Tue, 28 Jun 2022 18:00:03 +0000 (14:00 -0400)]
[AMDGPU] gfx11 WMMA instruction support

gfx11 introduces new WMMA (Wave Matrix Multiply-accumulate)
instructions.

Reviewed By: arsenm, #amdgpu

Differential Revision: https://reviews.llvm.org/D128756

2 years ago[flang][NFC] Fix warning
Valentin Clement [Thu, 30 Jun 2022 14:56:29 +0000 (16:56 +0200)]
[flang][NFC] Fix warning

2 years ago[pseudo] Forest dump ascii art isn't broken by large indices
Sam McCall [Thu, 30 Jun 2022 14:52:55 +0000 (16:52 +0200)]
[pseudo] Forest dump ascii art isn't broken by large indices

2 years ago[libc++] Remove dead code and unneeded C++03 specializations from type_traits
Nikolas Klauser [Thu, 30 Jun 2022 12:11:26 +0000 (14:11 +0200)]
[libc++] Remove dead code and unneeded C++03 specializations from type_traits

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D128906

2 years ago[libc++] Implement P0618R0 (Deprecating <codecvt>)
Nikolas Klauser [Thu, 30 Jun 2022 11:47:26 +0000 (13:47 +0200)]
[libc++] Implement P0618R0 (Deprecating <codecvt>)

Reviewed By: ldionne, #libc

Spies: cfe-commits, llvm-commits, libcxx-commits

Differential Revision: https://reviews.llvm.org/D127313